AI Tracks Colorectal Liver Metastases on Follow-up CT

Key Takeaways

• A retrospective single-center analysis of 87 adults with unresectable colorectal liver metastases used 174 contrast-enhanced CT scans obtained at baseline and about 8 weeks, a three-radiologist consensus reference, and a 52/35 training-test split to develop automated whole-liver lesion tracking.
• On the independent 35-pair test set, the affine plus Auto-MBT pipeline matched 458 of 464 lesions with matched-lesion F1 0.989 and 99% precision and recall, while detecting 28 of 30 new lesions with 90%/93% precision and recall and identifying 42 of 44 disappeared lesions with 93%/96% precision and recall.
• Against consensus, the affine plus Auto-MBT approach exceeded the listed comparator pipelines and individual radiologists across matched, new, and disappeared lesion categories, and its confidence-guided review signals tracked which discrepancy alerts were later accepted on adjudication.

Whole-liver follow-up in multifocal colorectal liver metastases are difficult to scale when lesions must be linked manually across serial CT studies. In a 35-pair internal test set, an automated affine registration plus model-based tracker matched 458 of 464 lesions and generated lesion-level confidence scores that the investigators positioned as a way to prioritize review.

At Memorial Sloan Kettering Cancer Center, Karunanayake et al. evaluated a retrospective single-center cohort of adults with unresectable disease who underwent baseline and first follow-up portal-venous-phase contrast-enhanced CT at about 8 weeks. Three abdominal radiologists created manual volumetric annotations and a consensus reference, and the workflow was built to match existing lesions while also flagging new lesions, disappeared lesions, and more complex split, merge, or other N-to-M evolution events.

Comparative testing showed the proposed pipeline outperforming deformable plus overlap and deformable plus Auto-MBT on internal validation, with matched-lesion F1 values of 0.797 and 0.959 for the comparator pipelines. Performance remained strong in scans containing more than 10 lesions and in subcentimeter lesions. On an external hepatic-lesion melanoma dataset, the same framework reached matched-lesion F1 0.994 across 23 scan pairs.

Against the consensus reference, the algorithm exceeded individual radiologists across matched, new, and disappeared lesion categories, with matched-lesion F1 0.989 (95% CI 0.975-0.995) versus 0.928-0.946, new-lesion F1 0.918 versus 0.507-0.585, and disappeared-lesion F1 0.944 versus 0.633-0.723. In blinded adjudication, 49 of 58 discrepancy alerts were accepted, and accepted alerts carried higher median confidence than rejected alerts, 0.720 versus 0.541, p=0.023; in high-tumor-count cases (>10 lesions/scan), the prespecified confidence bands showed accuracy/revision of 0.99/0.03 at ≥0.80, 0.97/0.18 at 0.60-0.80, and 0.92/0.23 at <0.60. The authors noted that the analysis was retrospective and single-institute, examined scan pairs rather than longer follow-up sequences, relied on manual lesion masks instead of end-to-end automated segmentation, was limited to hepatic lesions on contrast-enhanced CT, and included small counts in some subgroups.

Clinician Questions

How accurate was automated CT lesion tracking for colorectal liver metastases on the independent internal test set?

On the independent colorectal liver metastases test set, affine plus Auto-MBT matched 458 of 464 lesions with 99% precision and recall for matched lesions, achieved matched-lesion F1 0.989, detected 28 of 30 new lesions with 90%/93% precision and recall, and identified 42 of 44 disappeared lesions with 93%/96% precision and recall.

How did the Auto-MBT pipeline compare with radiologists and other lesion-tracking pipelines in colorectal liver metastases CT?

In colorectal liver metastases CT, matched-lesion F1 was 0.989 for affine plus Auto-MBT versus 0.797 for deformable plus overlap and 0.959 for deformable plus Auto-MBT, and against consensus the algorithm exceeded individual radiologists across matched, new, and disappeared lesion categories, including matched-lesion F1 0.989 (95% CI 0.975-0.995) versus 0.928-0.946 for radiologists.

What did lesion-level confidence scores show in colorectal liver metastases CT tracking?

In colorectal liver metastases CT tracking, 49 of 58 discrepancy alerts were accepted after blinded adjudication, accepted alerts had higher median confidence than rejected alerts at 0.720 versus 0.541 (p=0.023), and prespecified confidence bands showed accuracy/revision of 0.99/0.03 at ≥0.80, 0.97/0.18 at 0.60-0.80, and 0.92/0.23 at <0.60.