Artificial intelligence system of faster region-based convolutional neural network surpassing senior radiologists in evaluation of metastatic lymph nodes of rectal cancer
Rectal cancer is one of the most common gastrointestinal cancers worldwide, and while treatments have improved over the past two decades, lymph node (LN) metastasis—when cancer spreads to nearby lymph glands—remains a major driver of recurrence and poor outcomes. For doctors, detecting these metastatic nodes on magnetic resonance imaging (MRI) scans is a high-stakes challenge: it requires analyzing tiny details like node shape, boundary, and signal intensity, and even experienced radiologists can disagree on results. But what if artificial intelligence (AI) could do it faster, more consistently, and more accurately?
A 2019 study published in the Chinese Medical Journal, led by researchers from Qingdao University, Beihang University, and several major hospitals in China, tested exactly that. The team evaluated a deep learning tool called Faster Region-based Convolutional Neural Network (Faster R-CNN) to see if it could outperform senior radiologists in identifying metastatic LNs in rectal cancer patients—and how well it matched the “gold standard” of pathological diagnosis (tissue samples from surgery).
The Study: How They Tested Faster R-CNN
The researchers collected data from 414 rectal cancer patients (who had not received pre-op chemotherapy or radiation) from six clinical centers across China. They used Faster R-CNN to analyze pelvic MRI scans and compared its results to three key benchmarks:
- Radiologist diagnoses: Evaluations from experienced radiologists using standard MRI criteria.
- Pathologist diagnoses: The gold standard, using tissue samples from surgery to confirm LN spread.
- Recurrence-free survival (RFST): Follow-up data on how long patients lived without cancer returning (tracked for 36 months).
The team focused on two questions: How consistent is Faster R-CNN with radiologists and pathologists? and Can it predict long-term outcomes as well as the gold standard?
Key Results: AI Beats Radiologists—But Not Pathologists
The findings revealed three critical takeaways:
1. Faster R-CNN aligns closely with radiologists (and better with pathologists)
- Correlation in node count: The number of metastatic LNs detected by Faster R-CNN and radiologists was very consistent (a correlation score of 0.912, where 1.0 is perfect agreement).
- Better alignment with pathology: When compared to pathologists, Faster R-CNN had a stronger correlation (0.448) than radiologists (0.134). This means the AI’s results were closer to the “truth” of what the tissue samples showed.
2. Faster R-CNN is more consistent with pathologists for LN staging
LN staging (N0 = no spread, N1 = 1–3 nodes, N2 = 4+ nodes) guides treatment decisions. The team measured consistency using a kappa coefficient (a score for agreement between methods):
- Faster R-CNN vs. pathologists: 0.573 (moderate-to-good agreement).
- Radiologists vs. pathologists: 0.473 (fair agreement).
Faster R-CNN was especially better at identifying N2 patients (the most advanced stage): it correctly classified 42 N2 patients, while radiologists only got 26 right.
3. All methods predict survival—but pathology is still the best
All three methods (AI, radiologists, pathologists) showed the same trend: higher LN stages meant lower recurrence-free survival. But there was a gap for N2 patients:
- Pathologists predicted a 65% 3-year recurrence-free survival rate for N2 patients.
- Faster R-CNN and radiologists both predicted an 85% rate.
While Faster R-CNN didn’t match pathology’s accuracy here, it still outperformed radiologists in linking staging to survival.
Why This Matters for Patients and Doctors
For doctors, the biggest advantage of Faster R-CNN is speed and consistency:
- Speed: It analyzes an MRI scan in 20 seconds—vs. 10 minutes for a radiologist.
- Consistency: AI doesn’t get tired or miss small details, and it uses the same criteria every time. This reduces variability between radiologists and helps doctors make faster, more reliable treatment decisions (e.g., whether to use pre-op chemo to shrink the cancer).
For patients, this could mean:
- Faster diagnoses: Less waiting for scan results.
- More personalized treatment: Better LN staging helps doctors tailor therapies to individual needs.
Limitations and Next Steps
Faster R-CNN isn’t perfect. Pathology remains the gold standard because it uses actual tissue samples from surgery—something AI (and radiologists) can’t do pre-op. The AI was also trained on radiologist-marked data, not pathology, which explains why it’s not as accurate as tissue testing.
That said, the study shows AI is already a powerful pre-op tool. As the technology improves—especially if trained on more pathology data—it could become even more valuable for early detection and staging.
The Bottom Line
Faster R-CNN surpasses senior radiologists in evaluating pelvic metastatic LNs for rectal cancer. It’s faster, more consistent, and closer to the gold standard of pathology than human readers. While it’s not a replacement for pathologists, it’s a game-changer for pre-op care—helping doctors make better decisions faster, and giving patients a better shot at successful treatment.
Trial Registration: www.chictr.org.cn (No. ChiCTR-DDD-17013842)
Original Study Citation: Ding L, Liu GW, Zhao BC, Zhou YP, Li S, Zhang ZD, Guo YT, Li AQ, Lu Y, Yao HW, Yuan WT, Wang GY, Zhang DL, Wang L. Artificial intelligence system of faster region-based convolutional neural network surpassing senior radiologists in evaluation of metastatic lymph nodes of rectal cancer. Chin Med J 2019;132:379–387. doi:10.1097/CM9.0000000000000095
Was this helpful?
0 / 0