From One-Disease-Per-Model to One Model for 146 Findings: Alibaba DAMO Academy's DAMO RADAR in Science, and the Open-Source Route to Universal Medical Imaging AI
Abdominal CT is widely regarded as the hardest CT to read. A single contrast-enhanced exam can contain hundreds of slices, with dozens of soft-tissue organs — liver, pancreas, gallbladder, kidneys, spleen, and bowel — squeezed together, their densities nearly identical and their lesion signals subtle. Even a mid-to-senior radiologist needs twenty minutes to half an hour to complete one report. Today, that clinical Achilles’ heel has met its challenger — from a Chinese team.
On September 18, 2026, DAMO RADAR, a generalist medical-imaging AI model built by Alibaba DAMO Academy jointly with the First Affiliated Hospital of Zhejiang University School of Medicine and collaborating institutions, was published in the journal Science (science.org). It is the world’s first expert-level generalist medical-imaging AI for abdominal CT: a single model covering 18 abdominal anatomical structures and identifying more than 146 conditions and radiological findings at once, with the model, code, and full technical framework open-sourced (People’s Daily).
This result matters beyond a list of impressive AUC numbers in a top journal. It signals a fundamental shift in the technical paradigm of medical-imaging AI — from a “one model, one disease” specialist path toward a “one model reads a whole anatomical region” generalist path. As DAMO Academy puts it, “the core value of a specialist model is high precision; the core value of a generalist model is broad coverage. The two routes are not about replacing each other but about playing to their strengths in different clinical scenarios” (ScienceNet).
Why Generalization Is the Hard Core of Medical-Imaging AI
To understand the significance of DAMO RADAR, you must first see the industry-wide bottleneck. For years, medical-imaging AI has developed almost exclusively along the “one disease, one model” path: one model for lung nodules, one for pancreatic cancer, one for diabetic retinopathy. Such specialist models often achieve excellent results on a single disease, but at a heavy cost.
Specialist models have three fatal weaknesses. The first is an unsustainable development cycle. Zhang Jianpeng, a senior algorithm expert at DAMO Academy, put it bluntly: one disease typically takes two to three years from project start to deployment, and “with thousands of human diseases, we could work a lifetime and never finish” (36Kr). The second is extreme reliance on manual annotation. To learn to recognize a lesion, an AI needs radiologists to draw and label abnormalities slice by slice. Radiologist time is expensive, labeled CT datasets are inherently small, and every model carries high training cost and long lead time. The third is poor generalization. A specialist model learns a “fixed exam type”; face it with a different disease or a different scanner, and it frequently fails.
More fundamentally, the real clinical world does not operate by “single disease.” A patient lying down for an abdominal CT may simultaneously harbor a liver mass, a subtle pancreatic density change, and a kidney stone. A doctor’s job is to read the whole volume and screen multiple organs at once — but a specialist AI can only answer the narrow question “does this patient have that one disease?” This is exactly the gap a generalist model fills.
DAMO RADAR breaks through by first recognizing anatomical structures, then linking them to pathology. It mimics how real radiologists read: lock onto the liver, pancreas, and biliary tract one by one, then examine each organ for abnormalities. Borrowing this logic, the team created the “organ-level fine-grained alignment” technique, an international first (Hangzhou Net).
The DAMO RADAR Architecture
DAMO RADAR stands for Rapid Abdominal Diagnosis with AI and Radiology. Its core methodology is vision-language learning. The team’s insight: in a hospital, every CT exam naturally comes with a diagnostic report written by a radiologist, describing in detail the organs and lesions. Rather than asking doctors to hand-annotate every lesion, why not let the model learn the inherent relationship between massive numbers of “image-report” pairs, so it can infer for itself what kind of imaging pattern corresponds to what disease?
But simply applying off-the-shelf vision-language learning to CT does not work. The reason is that CT signals are sparse — the fraction of genuinely meaningful information in a 3D volume is tiny, so naive image-text alignment yields poor accuracy. The team countered with two innovations:
- Organ-level fine-grained alignment: decompose the 3D CT volume into independent “anatomical units,” and align image and report text precisely at the organ level. The model first learns where the liver, pancreas, and gastrointestinal tract are, then links each organ’s imaging pattern to the report’s description of that organ, filtering out the interference of vast amounts of normal background tissue.
- Adaptive contrastive modeling: conventional contrastive learning treats any two different samples as “unrelated,” but medically, two patients with healthy livers should be regarded as similar. Adaptive contrastive modeling adjusts the relationship between samples based on medical knowledge, more accurately modeling the boundary between rare findings and normal appearance.
At the implementation level, DAMO RADAR combines a 3D vision branch that ingests CT volumes with BERT-family text encoders in both Chinese and English, connected through a contrastive learning objective (navsplace). It also outputs attention maps, letting clinicians see exactly which region, slice, and feature drove a prediction — a concrete interpretability and safety feature that pure text-based medical chatbots lack.
For training, the team built RAD-CT, a dedicated dataset of 424,911 contrast-enhanced abdominal CT exams, yielding roughly 15 million anatomy-aware image-text pairs. Because the model learned directly from existing clinical reports rather than hand-drawn annotations, a dataset of this scale became practical to assemble (navsplace).
Figure 1: End-to-end DAMO RADAR architecture (input → multi-organ localization → multi-disease classification)
DAMO RADAR End-to-End Pipeline
+--------------------------------------------------------------+
| Contrast-Enhanced Abdominal CT Volume Input |
| (hundreds of axial slices, tens of organs) |
+-------------------------------+------------------------------+
|
v
+--------------------------------------------------------------+
| 1. 3D Vision Branch (3D Vision Encoder) |
| - encode CT volume into 3D features |
| - decompose volume into anatomical "units" (organ) |
+-------------------------------+------------------------------+
| organ-level features
v
+--------------------------------------------------------------+
| 2. Bilingual Text Branch (BERT Encoder) |
| - Chinese report encoding ----+ |
| - English report encoding ----+--- image-text contrast |
+-------------------------------+------------------------------+
| alignment
v
+--------------------------------------------------------------+
| 3. Multi-disease Head (146 findings / 18 structures) |
| - organ-level lesion box (Liver/Pancreas/Kidney/GB..) |
| - finding-level class (tumor/inflam/bleed/obstruct..) |
| - attention map (interpretability) |
+-------------------------------+------------------------------+
|
v
+--------------------------------------------------------------+
| Output: structured diagnostic hints + region-level heatmaps |
| -> for radiologist review and clinical decision |
+--------------------------------------------------------------+
Figure 2: “One-disease-per-model” vs DAMO RADAR generalist
Specialist "One-Disease-Per-Model" DAMO RADAR Generalist
+--------------------------+ +-----------------------------+
| Lung-nodule AI (special)| | |
| Pancreatic-cancer AI | | One model reads the whole |
| Gastric-cancer AI | | abdomen |
| Colorectal-cancer AI | | (18 structures / 146 texts) |
| Retinopathy AI | | |
| ... one model per disease| | sync screening, one pass |
+--------------------------+ +-----------------------------+
high precision (one disease) broad coverage (many)
needs heavy manual annotation learns from clinical reports
2-3 yr dev cycle per disease no extra manual annotation
poor generalization migrates to emergency too
Multi-Center Clinical Validation: A Rigorous Physical Exam
Papers can be polished, but numbers do not lie. To prove DAMO RADAR is not a lab “demo,” the team designed a punishing multi-layer validation program spanning internal real-world cohorts, external multi-center data, pathology gold standards, and reader studies.
First, the internal real-world cohort. Across nearly 39,000 real-world consecutive contrast-enhanced abdominal CT exams covering 146 findings, DAMO RADAR achieved a mean AUC of 0.913 (confidence interval 0.911-0.915). AUC measures a model’s ability to distinguish “diseased” from “healthy”; 1.0 is a perfect discriminator, 0.5 is a coin flip, and anything above 0.9 is considered excellent (People’s Daily).
The real test was external generalization. On more than 24,000 CT exams from 8 external hospitals spanning different regions and scanner vendors, AUC ranged from 0.874 to 0.912; even on a cross-population cohort without fine-tuning, AUC remained 0.883 (navsplace).
The hardest check was against the pathological gold standard. For hepatocellular carcinoma, pancreatic cancer, gastric cancer, and colorectal cancer, the team benchmarked the model against biopsy-confirmed gold standards, achieving AUCs of 0.891 to 0.984. This indicates the model’s tumor detection is not a coarse guess but highly consistent with pathology.
DAMO RADAR also withstood the emergency department test. Emergency differs sharply from routine outpatient scanning — time is short, contrast may not peak, patients may not cooperate, and image quality suffers. Critically, emergency was outside the training objective. When the team tested roughly 27,000 emergency cases, the model still reached AUC 0.904 on acute abdomen findings, demonstrating native generalization (36Kr).
Figure 3: Multi-layer clinical validation workflow
DAMO RADAR Multi-Layer Clinical Validation
+---------------------------------------------------------------+
| 1. Internal real-world cohort |
| ~39k exams / 146 findings / 18 structures |
| -> mean AUC = 0.913 (0.911-0.915) |
+-------------------------------+-------------------------------+
|
+-------------------------------v-------------------------------+
| 2. External multi-center generalization |
| 8 external centers / >24k exams / multi-region devices |
| -> AUC 0.874-0.912 ; cross-population no-finetune 0.883 |
+-------------------------------+-------------------------------+
|
+-------------------------------v-------------------------------+
| 3. Pathology gold-standard comparison |
| liver/pancreatic/gastric/colorectal -> AUC 0.891-0.984 |
+-------------------------------+-------------------------------+
|
+-------------------------------v-------------------------------+
| 4. Emergency generalization (outside training) |
| ~27k acute-abdomen cases -> AUC 0.904 |
+-------------------------------+-------------------------------+
|
+-------------------------------v-------------------------------+
| 5. Reader study (human-AI) |
| 14 centers / 26 radiologists / 300 patients |
| model beat 23/26; AI -> sensitivity +10%, time -30.7% |
+---------------------------------------------------------------+
The reader study is the direct evidence of expert-level performance. The team recruited 26 radiologists from 14 institutions, including 11 senior and 15 junior physicians, and had both doctors and model independently interpret CT images from 300 patients. DAMO RADAR outperformed 23 of the 26 radiologists on average; only 3 senior radiologists edged slightly higher (ScienceNet).
Yet DAMO RADAR’s design philosophy is never to replace the doctor. In human-AI collaborative reading tests, with AI hints, radiologists improved their detection sensitivity (leak-prevention ability) by about 10% and cut per-case reading time by 30.7%, and in some tests junior radiologists assisted by AI surpassed the sensitivity of senior radiologists (ScienceNet). This points straight at the most compelling value of a generalist medical-imaging AI: rapidly lifting junior radiologists to a senior expert’s diagnostic level — precisely what grassroots hospitals need most.
Engineering Details Hidden Inside the Medical Detection Architecture
From an engineering angle, DAMO RADAR’s “generality” is not bought by brute-forcing more data but by a deliberately designed three-stage medical detection pipeline: anatomical localization → per-organ classification → finding aggregation. Let us unpack it as a runnable diagnostic pipeline.
At the multi-organ anatomical-structure level, the model does not make an end-to-end disease judgment directly on the whole CT. First it performs organ-level localization — boxing out the liver, pancreas, kidneys, gallbladder, and other anatomical structures in the 3D volume, establishing each organ’s “territory.” This is like delimiting the boundary of the problem domain before making fine judgments inside it. It transforms the ill-posed problem of “finding needles in hundreds of slices” into a well-posed task of “checking for abnormalities within known organ regions,” sharply reducing false positives and misses.
At the multi-disease classification level, each localized organ region is fed to a multi-task classification head that simultaneously outputs organ-level findings (e.g., “liver space-occupying lesion”), imaging-sign level manifestations (e.g., “ascites,” “bowel obstruction”), and confidence. Finally, these multi-level outputs are aggregated into a “structured diagnostic hint” telling the doctor “which organ, which region, likely what problem, and on what evidence (attention map).”
Figure 4: One model, many diseases — organ localization to multi-classification decision flow
Single CT volume -> one-model-many-disease decision tree
+--------------------------------------------------------------+
| 3D CT volume |
+---------------------------+----------------------------------+
| organ localization
+------------------+------------------+
v v v
+----------+ +------------+ +------------+
| Liver | | Pancreas | | Kidney | ...
+----+-----+ +-----+------+ +-----+------+
| multi-task | multi-task | multi-task
v v v
+------------------------------------------------------+
| Multi-disease head (146 findings) |
| +- organ-level: liver mass/pancreas change/stone |
| +- sign-level: ascites/bowel obstruction/AAA |
| +- malignancy: liver/pancreas/gastric/colorectal |
| +- confidence + attention map |
+------------------------------------------------------+
|
v
structured diagnostic hint for doctor review
To ground this in engineering, here is a simplified Python skeleton of a “multi-task abdominal CT pipeline” reflecting the “organ localization → per-organ classification → finding aggregation” idea. The real model processes 3D volumes end-to-end; this uses 2D slices merely to illustrate the multi-task layering:
# multi_task_abdominal_pipeline.py
import numpy as np
from typing import Dict, Tuple
FINDING = {
"liver": ["focal_lesion", "cirrhosis", "normal"],
"pancreas": ["mass", "necrosis", "dilated_duct", "normal"],
"kidney": ["stone", "cyst", "hydronephrosis", "normal"],
}
def organ_localization(vol: np.ndarray) -> Dict[str, Tuple[int, int]]:
"""DAMO RADAR-style: decompose 3D volume into anatomical units."""
organs = list(FINDING.keys())
n = vol.shape[0]
step = n // len(organs)
return {o: (i * step, (i + 1) * step) for i, o in enumerate(organs)}
def classify_region(region: Tuple[int, int]) -> Dict[str, float]:
"""Per-region multi-task head over 18 structures."""
# demo logit probabilities; real values come from the 3D network
return {f: float(np.random.rand()) for f in FINDING.keys()}
def run_inference(vol: np.ndarray) -> Dict[str, Dict[str, float]]:
boxes = organ_localization(vol)
return {o: classify_region(r) for o, r in boxes.items()}
if __name__ == "__main__":
fake = np.random.rand(150, 256, 256) # 150 CT slices
for organ, _ in run_inference(fake).items():
print(f"scan organ: {organ}")
This is only an illustrative skeleton — the real DAMO RADAR is an end-to-end 3D vision-language model, not an independent slice-by-slice classifier — but it captures the core layered pattern of “anatomical-unit splitting + multi-task classification + finding aggregation” that grounds “generality” in engineering.
These clinical validation metrics (mean AUC 0.913, +10% sensitivity, -30.7% reading time) do not materialize by themselves. In practice they are produced by a rigorous metric-computation pipeline. Below is a Python sketch for “multi-center clinical validation metrics,” covering AUC, sensitivity, specificity, and bootstrap confidence intervals — the standard workflow for quantitative reporting:
# validation_metrics.py
import numpy as np
from sklearn.metrics import roc_auc_score
def bootstrap_ci(y_true, y_score, n_boot=1000, seed=42):
rng = np.random.default_rng(seed)
aucs = []
idx = np.arange(len(y_true))
for _ in range(n_boot):
s = rng.choice(idx, size=len(idx), replace=True)
if len(np.unique(y_true[s])) < 2:
continue
aucs.append(roc_auc_score(y_true[s], y_score[s]))
lo, hi = np.percentile(aucs, [2.5, 97.5])
return float(np.mean(aucs)), lo, hi
def sensitivity_specificity(y_true, y_score, threshold=0.5):
y_pred = (y_score >= threshold).astype(int)
tp = np.sum((y_pred == 1) & (y_true == 1))
fn = np.sum((y_pred == 0) & (y_true == 1))
tn = np.sum((y_pred == 0) & (y_true == 0))
fp = np.sum((y_pred == 1) & (y_true == 0))
sens = tp / (tp + fn) if (tp + fn) else 0.0
spec = tn / (tn + fp) if (tn + fp) else 0.0
return sens, spec
if __name__ == "__main__":
rng = np.random.default_rng(0)
y_true = rng.integers(0, 2, size=2000)
y_score = rng.uniform(0, 1, size=2000)
auc, lo, hi = bootstrap_ci(y_true, y_score)
sens, spec = sensitivity_specificity(y_true, y_score)
print(f"mean AUC={auc:.3f} 95%CI=[{lo:.3f},{hi:.3f}] "
f"sens={sens:.3f} spec={spec:.3f}")
Bilingual Paths and Open-Source Deployment Inference
In a real system, DAMO RADAR routes Chinese and English reports through separate BERT encoders and is deployed service-style on the radiology PACS side. Here is a sketch of its inference service HTTP entry logic (Python):
# radar_service.py
import numpy as np
class RadarService:
def __init__(self, detector, bilingual=True):
self.detector = detector
self.bilingual = bilingual
def diagnose(self, volume: np.ndarray, lang="zh"):
findings = self.detector(volume) # (region, scores) list
out = []
for organ, score_map in findings:
top = max(score_map, key=score_map.get)
out.append({"organ": organ,
"top_finding": top,
"top_score": round(score_map[top], 4),
"lang": lang})
return {"finding_count": len(out), "items": out}
Figure 5: Clinical workflow integration (human-AI collaborative loop)
Abdominal CT Clinical Workflow Integration (Human-AI Loop)
+--------------+ +----------------------+ +----------------+
| CT scanning | ----> | images auto-upload | ----> | DAMO RADAR |
| acquisition | | to PACS | | AI side engine |
+--------------+ +----------------------+ | (parallel) |
+--------------+ +----------------------+ +-------+--------+
| raw slices | <---- | radiologist console | <------------+
+--------------+ | - AI structural hints| - detect 18/146
| - stacked attention map | - hints + heatmap
| - doctor confirm/override |
+----------+--------------+
| final sign-off
v
diagnostic report
Open-Source Inclusiveness: Technology Not Locked Behind High Walls
If technical quality determines how far DAMO RADAR can go, its licensing strategy determines how many people can benefit. Here the team made a decision of far-reaching significance to global medicine: model weights, core code, and the full technical framework are fully open-sourced (Hangzhou Net).
Concretely, the release follows a common “code vs weights layered licensing” practice: code on GitHub under Apache 2.0, so research teams can freely study, modify, and extend the pipeline; model weights on Hugging Face under CC BY-NC-SA 4.0 (non-commercial, share-alike). Read together, the implication is clear: a hospital IT team can freely study the pipeline, but anyone planning a commercial product on top of the released weights needs a separate arrangement with DAMO Academy (navsplace). This is a common safety gate in open-source medical AI, preventing algorithms from being indiscriminately leveraged into paid diagnostics.
Per public documentation, inference can run on a single NVIDIA A100 or H20 data-center GPU, while multi-GPU accelerates large-scale work; reproducing training from scratch is far beyond consumer hardware — the developers used 24 such GPUs (IntelligentLiving).
Figure 6: Open-source deployment topology (GitHub / Hugging Face → global medical research institutions)
DAMO RADAR Open-Source Deployment Topology
+-------------------------------------+
| Alibaba DAMO Academy |
| release - maintain - update |
+-------+-----------------------------+
| open source
+----------------------------+---------------+
v v
+-------------------------+ +-----------------------+
| GitHub (Apache 2.0) | | Hugging Face |
| - code / fine-tune | | (CC BY-NC-SA 4.0) |
| - inference / docs | | - model checkpoints |
+------------+------------+ +-----------+-----------+
| free for global research/medical use |
v v
+-----------------------------------------------+
| local deployment (single A100/H20 suffices) |
| - PACS integration - local eval - fine-tune |
| - research reproduction - independent audit |
+-----------------------------------------------+
This open-source decision reflects a deep divergence between Chinese and American AI development routes. Stanford’s AI Index 2026 reports that 84% of Chinese respondents hold a positive attitude toward AI, versus only 38% in the United States (Toutiao copy of Yuedu). Many analysts attribute public optimism to the inclusive route of technology deployment. Medical cost differences magnify the route gap concretely: a contrast-enhanced CT at a Chinese public hospital costs about 200 RMB, while a comparable US CT bill can reach $7,000, a heavy burden for ordinary patients (Toutiao copy of Yuedu).
Figure 7: China vs US medical-AI route comparison
China vs US Medical-AI Route Comparison
+----------------------------------+----------------------------------+
| China route | US route |
+----------------------------------+----------------------------------+
| emphasizes inclusive open access | leading firms favor closed |
| flagships open-source model+code | pay-per-use, technology moats |
| global research can reuse | restricted commercial licensing |
+----------------------------------+----------------------------------+
| public CT: ~200 RMB | comparable CT: ~$7,000 |
+----------------------------------+----------------------------------+
| 84% positive toward AI | 38% positive toward AI |
| (Stanford AI Index 2026) | (Stanford AI Index 2026) |
+----------------------------------+----------------------------------+
AI Index data: per Toutiao copy of Yuedu compilation
For grassroots hospitals, this open-source strategy is timely relief. They see limited reading volume and relatively inexperienced staff, and abdominal CT is the hardest to read there — exactly where a generalist model acting as a “tireless auxiliary eye” can fill the gap (Hangzhou Net). Letting advanced technology reach more ordinary people, releasing cutting-edge results from behind high walls, is the very essence of inclusive medical AI.
Below is a Python inference sketch showing how to load RADAR-style weights and run universal diagnosis on a local CT volume (illustrative interface; use the official repo for production):
# radar_inference_demo.py
import numpy as np
def run(ct_volume: np.ndarray, region_box: tuple, encoders, tokenizer):
"""Run RADAR-style universal diagnosis on a local CT volume."""
z0, z1, y0, y1, x0, x1 = region_box # organ bounding box
patch = ct_volume[z0:z1, y0:y1, x0:x1]
visual = encoders["vision"](patch[None])
probs = {}
for finding in ["liver_mass", "pancreatic_ca", "renal_stone",
"cholecystitis", "bowel_obstruction"]:
text = encoders["text"](tokenizer(finding, return_tensors="pt"))
probs[finding] = float((visual * text).sum().sigmoid())
return sorted(probs.items(), key=lambda kv: -kv[1])
Two components power the training regime behind RAD-CT and the organ-level alignment. First, an adaptive contrastive loss that treats organ-aware pairs as positives; second, an organ-level alignment objective that decomposes each volume into anatomical units and matches them to report segments. A compact implementation sketch:
# train_organ_alignment.py
import torch
import torch.nn as nn
class AdaptContrastiveLoss(nn.Module):
"""Adaptive contrastive loss with organ-aware positives."""
def __init__(self, temperature=0.07):
super().__init__()
self.temperature = temperature
def forward(self, vis, txt, mask):
# vis, txt: (B, D) features; mask: organ-aware positive matrix
logits = torch.matmul(vis, txt.T) / self.temperature
logits = logits - logits.max(dim=-1, keepdim=True).values
exp = torch.exp(logits)
denom = exp.sum(dim=-1, keepdim=True)
pos = (exp * mask).sum(dim=-1) / (mask.sum(dim=-1) + 1e-9)
return -torch.log(pos / (denom + 1e-9)).mean()
def train_one_epoch(model, opt, loader, crit, device):
model.train()
total = 0.0
for vis, txt, mask in loader:
vis, txt, mask = vis.to(device), txt.to(device), mask.to(device)
v = model.vision_branch(vis)
t = model.text_branch(txt)
loss = crit(v, t, mask)
opt.zero_grad()
loss.backward()
opt.step()
total += loss.item()
return total / len(loader)
For model debugging and radiology trust, extracting per-volume attention heatmaps is essential. Here is a sketch that turns the vision transformer’s attention into a 2D, up-sampled, normalized heatmap matching the raw slice grid:
# attention_map.py
import torch
import torch.nn.functional as F
def extract_attention(vit, volume: torch.Tensor, layer=-1, patch=16):
feats, attn = vit(volume, output_attentions=True)
heads = attn[layer] # (B, H, N, N)
cls = heads[:, :, :, 0].mean(dim=1) # CLS -> all patches
cls = cls[0, 1:] # drop CLS token
n = int(cls.numel() ** 0.5)
cls = cls.reshape(n, n) # 2D attention map
cls = (cls - cls.min()) / (cls.max() - cls.min() + 1e-9)
return F.interpolate(cls[None, None], scale_factor=patch,
mode="bilinear")
To build RAD-CT’s anatomy-aware image-text pairs at scale, the pipeline pairs each 3D volume with report sentences that are organ-anchored. A schema-level sketch in Python:
# radct_pair_builder.py
from dataclasses import dataclass, field
from typing import Dict, List
@dataclass
class RadPair:
volume_id: str
organ: str
report_sentence: str
snippet_range: tuple
tags: List[str] = field(default_factory=list)
ORGAN_ANCHORS = {
"liver": ["liver", "hepatic", "segment viii"],
"pancreas": ["pancreas", "uncinate", "tail"],
"kidney": ["right kidney", "left kidney", "renal"],
}
def extract_pairs(volume_id: str, report: str) -> List[RadPair]:
pairs = []
for organ, anchors in ORGAN_ANCHORS.items():
for anchor in anchors:
if anchor in report.lower():
pairs.append(RadPair(volume_id, organ,
report, (0, 64), [organ]))
break
return pairs
At the hospital IT level, radiology departments often use Go to build the middleware that talks to PACS, pulls DICOM-compliant CT series, selects the enhancement phase, and hands them to the AI inference service. Below is an illustrative Go imaging queue dispatcher sketching how “AI-assisted reading” is plugged in hospital-side:
// pacs_dispatcher.go
package main
import (
"context"
"log"
"sync"
)
type Volume struct {
PatientID string
Series []string // DICOM image URL list
}
type RadarClient interface {
Diagnose(ctx context.Context, vol *Volume) (string, error)
}
type Dispatcher struct {
client RadarClient
workerCount int
}
func (d *Dispatcher) Run(ctx context.Context, jobs <-chan *Volume) {
var wg sync.WaitGroup
for i := 0; i < d.workerCount; i++ {
wg.Add(1)
go func() {
defer wg.Done()
for v := range jobs {
out, err := d.client.Diagnose(ctx, v)
if err != nil {
log.Printf("diagnose %s: %v", v.PatientID, err)
continue
}
log.Printf("hint for %s: %.30s", v.PatientID, out)
}
}()
}
wg.Wait()
}
func main() {
ctx := context.Background()
jobs := make(chan *Volume, 64)
disp := &Dispatcher{client: nil, workerCount: 4}
go func() {
jobs <- &Volume{PatientID: "P1001"}
close(jobs)
}()
disp.Run(ctx, jobs)
}
The above Go snippet is an illustrative concurrency pattern demonstrating the “PACS fetch → AI diagnose → hint back” integration frame; rely on the official open-source repo for production.
Because DAMO RADAR is trained on the portal-venous phase of contrast-enhanced CT, the hospital middleware must filter the correct DICOM series before inference. A small Go helper for phase selection:
// phase_select.go
package main
import (
"strings"
)
type Series struct {
UID string
Name string
Phase string // "portal-venous", "arterial", "non-contrast"
}
var phaseKeys = []string{"portal venous", "portal", "venous", "PV"}
func SelectPortal(series []Series) *Series {
for i := range series {
lower := strings.ToLower(series[i].Name + series[i].Phase)
for _, k := range phaseKeys {
if strings.Contains(lower, k) {
return &series[i]
}
}
}
return nil
}
func Validate(vol [][][]int16) bool {
// ensure 3D volume shape is a multiple of the patch grid
if len(vol) == 0 || len(vol[0]) == 0 || len(vol[0][0]) == 0 {
return false
}
return true
}
When a research group wants to validate on local data before trusting the paper’s numbers, a fine-tuning loop on a small internal cohort is the standard first step. A compact PyTorch fine-tune driver:
# finetune_local.py
import torch
from torch.utils.data import DataLoader
def finetune(model, dataset, epochs=10, lr=1e-5, device="cuda"):
model.to(device)
opt = torch.optim.AdamW(model.parameters(), lr=lr)
loader = DataLoader(dataset, batch_size=8, shuffle=True)
for ep in range(epochs):
total_loss, n_batch = 0.0, 0
for vol, report, organ in loader:
vol = vol.to(device)
loss = model.forward(vol, report) # organ-aware objective
opt.zero_grad()
loss.backward()
opt.step()
total_loss += loss.item()
n_batch += 1
print(f"epoch {ep}: loss={total_loss / n_batch:.4f}")
return model
From Abdominal CT to Universal Medical Imaging: The Road Ahead
DAMO RADAR currently focuses on the abdomen, but its ambition goes further. The team states clearly that this research paradigm — organ-level fine-grained alignment plus adaptive contrastive modeling plus vision-language learning — not only serves multi-disease recognition in abdominal CT but can migrate to other forms of medical imaging, accelerating the arrival of artificial general intelligence (People’s Daily).
The generalization path is technically clear. Prior open-source radiology generalists like RadFM and Stanford’s CheXagent either cover different modalities or focus on chest X-rays; nobody had yet cracked abdominal CT, whose structure is more complex, soft tissues more numerous, and densities closer. DAMO RADAR enters a relatively open territory (navsplace). Once the “anatomical-unit-level alignment” methodology is proven, it can naturally migrate to the chest (lung, mediastinum, heart), pelvis, head and neck, and even to MRI, ultrasound, and PET.
Deeper still is the replicability of the methodology. DAMO RADAR proves the feasibility of “learning universal diagnosis directly from the massive ‘image-report’ pairs hospitals already possess.” It bypasses the manual-annotation bottleneck that has slowed medical-imaging AI for a decade — by learning directly from the free-text reports radiologists already produce, the model absorbs hundreds of thousands of real clinical knowledge points with zero annotation. That is precisely why one model can cover 146 findings instead of 146 separate models (navsplace). It also has operational consequences: hospitals evaluating this stack move from managing a portfolio of single-disease tools to deploying one generalist with a consistent interface, attention maps for interpretability, and a single update path.
Figure 8: Future generalization path (abdominal CT → multi-region / multi-modal)
DAMO RADAR Research-Paradigm Generalization Path
+------------------+ anatomical-unit alignment + vision-language + contrast
| Abdominal CT | ---------------------------------------------------+
| (validated) | |
+------------------+ |
18 structures / 146 findings |
AUC 0.913 |
| v
+------------------> 1. Chest CT (lung/mediastinum/heart)
+------------------> 2. Pelvic CT (uterus/prostate/bladder)
+------------------> 3. Head & neck CT (brain/soft tissue)
+------------------> 4. MRI (finer soft-tissue contrast)
+------------------> 5. Ultrasound (real-time / no ionizing)
+-------------------------------------------------------------+
| Endpoint: multi-region, multi-modal universal medical- |
| imaging AI (pushing the healthcare industry toward AGI) |
+-------------------------------------------------------------+
We must also be clear-eyed about the boundaries of any top-journal paper. The researchers acknowledge several limitations: the sensitivity-improvement figure does not specify absolute versus relative improvement; the human-comparison was a “Reader Study” framed by the authors themselves, not a true randomized clinical trial, and no patient outcomes were measured; third-party reproduction is still pending (navsplace). Moreover, as of September 2026 the model has not received clinical approval and is released as a research model for study, validation, and further development (IntelligentLiving). These boundaries remind the field that a page of expert-level numbers still has a serious road to travel before regulatory approval and bedside deployment.
Conclusion: The Radar of Generalist Medical-Imaging AI Is On
Back to the opening question: why abdominal CT? Because it is among the hardest. Why DAMO RADAR? Because it achieves expert-level universal recognition on the hardest CT. When DAMO Academy consolidates three years of work — five Nature Medicine papers and, now, this Science publication — across pancreatic, gastric, and colorectal cancer into a single generalist model that reads the entire abdomen, it rewrites the narrative of medical-imaging AI: the era of one-model-per-disease radiology AI is ending faster than most hospital procurement cycles can keep up with (navsplace).
More importantly, this is a double victory of both technology and route. Technologically, it uses organ-level fine-grained alignment to crack the problem of general modeling under sparse CT signals. In terms of route, it uses full open source to practice the value of inclusive medical AI. When a medical model capable of identifying 146 diseases at expert level is opened to the world, code and weights together, it pushes outward not just diagnostic accuracy but a conviction: advanced technology should leave the high walls and help light the way toward better health for more ordinary people.
The next milestones to watch are whether independent multi-center validation can reproduce the paper’s numbers, whether regulatory decisions let the model reach the bedside, and whether grassroots hospitals can actually use this “tireless auxiliary eye.” One way or another, the generalist door of medical-imaging AI has been pushed open by this result from a Hangzhou laboratory — and it is not going to close.