AI Science Tracker
EN

National Cancer Institute

2 discoveries from 2021 to 2024

Discoveries

Automated prostate-extension grading shows a sensitivity–specificity tradeoff against pathology

Source check: Source checked. Not independent scientific confirmation.

Finding: A cascaded segmentation and random-forest pipeline was evaluated against surgical pathology in a retrospective prostate-cancer cohort. In 127 held-out patients, one threshold produced lower sensitivity but higher specificity than the radiologist comparison. This is a bounded clinical-validation finding, without established superiority, improved surgical decisions or independent replication of the reused segmentation model.

Technology: Reused 3D U-Net/residual segmentation pipeline with random-forest extraprostatic-extension grading

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
Learned segmentation extracted lesion/capsule features and a trained random forest mapped them to extension grades; comparison with actual surgical pathology supplies the distinct clinical-validation result. This extends a previously developed segmentation workflow rather than constituting independent replication of it.
Test and result
The source split 634 prostatectomy patients into 507 development and 127 test patients. Fixed segmentation models produced prostate/lesion masks; geometric features trained random forests with repeated cross-validation. At EPE≥1, test sensitivity/specificity were 0.67/0.73 versus 0.81/0.62 for the radiologist, using pathology as reference. Four-grade balanced accuracy was limited, and individual grade AUCs ranged 0.55–0.70.
Limitation / still unknown
This retrospective analysis used prospectively recorded radiologist assessments, without prospective AI-model locking. Development grades partly incorporated pathology, while comparison grades came from the radiologist. Reused segmentation models and overlapping institutional recruitment periods mean the full pipeline is not established as wholly unfamiliar with test patients. No paired clinical reader-assistance or treatment-benefit trial was performed, and no general superiority is claimed.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC11490411/ · https://pmc.ncbi.nlm.nih.gov/articles/PMC10575564/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 21 Nov 2026. Review scope: Read actual primary clinical selection/patient split, radiologist/pathology label construction, fixed segmentation-feature and random-forest training/evaluation Methods, pathology-reference test Results and comparison limits. Source reference27 checked against previously read segmentation paper. Individual overlap across reused model training and this cohort not reconstructed. Exact original article front DOI/title and fresh PubMed primary relationship metadata checked..

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 1/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Cascaded prostate-MRI model shows limited lesion detection and classification in clinical testing

Source check: Source checked. Not independent scientific confirmation.

Finding: A cascaded 3D U-Net and residual-network system was evaluated on held-out clinical prostate MRI scans, with partial targeted-biopsy comparison. It detected 56.1% of expert-marked lesions and achieved 30.8% overall PI-RADS classification accuracy. The result documents limitations of retrospective clinical performance, without established cancer-screening benefit or evidence that the system can replace biopsy.

Technology: Cascaded 3D U-Net lesion segmentation and 3D residual networks for BPH filtering and PI-RADS classification

Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
Learned networks detected and classified suspicious regions from patient MRI; comparison with expert labels and available biopsy outcomes produced the represented diagnostic-validation result. Low performance is retained rather than treated as evidence of clinical utility.
Test and result
The source reports 150 held-out scans with 198 expert-marked lesions. After benign-hyperplasia filtering, the algorithm detected 111 of 198 lesions, with PPV 62.7% and overall PI-RADS accuracy 30.8%. Among 136 lesions with targeted biopsy, 82.3% of detected reference lesions were tumor-positive, while 50.9% of missed reference lesions were benign. This is a specific clinical evaluation with partial pathology follow-up, beyond a public image benchmark alone.
Limitation / still unknown
Detection/reference categories were defined by one expert radiologist rather than complete histopathology. False-positive regions lacked targeted-biopsy validation, preventing full cancer detection-rate assessment. Cohort totals are supported by Figure 1, but abstract/Methods split counts conflict; pathology-grade subtotals also do not fully reconcile. Prostate-X includes repeated scans, and individual-patient separation across all partitions was not reconstructed. No reader-assistance benefit, prospective workflow evaluation or clinical efficacy was demonstrated.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://pmc.ncbi.nlm.nih.gov/articles/PMC10575564/

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.

Review date: 21 Nov 2026. Review scope: Read actual primary clinical cohort, expert reference, cascaded 3D U-Net/residual architecture and hold-out evaluation Methods; test detection/classification, partial biopsy Results and limitations. Directly viewed Figure1; split-count and pathology-grade discrepancies preserved. Supplements and individual scans not independently reconstructed. Exact original article front DOI/title and fresh PubMed primary relationship metadata checked..

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 1/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open the interactive profile for the timeline, or browse all organizations and model families.