Histology Model for Homologous Recombination Deficiency

Evidence

Homologous recombination deficiency is presented as a biomarker relevant to treatment decisions in high-grade serous tubo-ovarian carcinoma. The investigators developed HRDPath, an end-to-end deep-learning architecture intended to predict homologous recombination deficiency status directly from hematoxylin and eosin-stained whole-slide images, without additional molecular testing. The study framed this image-based approach against genomics-based homologous recombination deficiency tests described as variable, expensive, time-consuming, and dependent on advanced genomic technologies. The reported task was biomarker classification from pathology images, not measurement of treatment response, survival, or patient benefit. (pmid:42907492)

The reported design was model development in an in-house UWOV cohort followed by independent validation in two unseen benchmarking datasets, TCGAOV and PTRC. UWOV contained 152 patients, including 145 with high-grade serous tubo-ovarian carcinoma and seven with other ovarian cancers. TCGAOV contained 81 patients with high-grade serous tubo-ovarian carcinoma and 82 diagnostic whole-slide images, while PTRC contained 93 patients with high-grade serous tubo-ovarian carcinoma. The development cohort was selected because it had multiple slides from different anatomical locations per patient; the independent datasets used diagnostic slides. This structure separates model construction from evaluation in datasets not used for development, although the supplied material does not provide all selection and labeling details needed to assess comparability across cohorts. (pmid:42907492)

UWOV included diagnostic whole-tissue-section slides from primary debulking resections. The source describes material from tubo-ovarian locations and extra-ovarian locations including omentum, abdomen, other pelvic sites, and other locations. HRDPath used weakly supervised learning, meaning it operated on whole-slide images without tumor segmentation. Its patient-level learning aggregated features across multiple slides from an individual patient, rather than treating every slide as an independent patient-level observation. Multi-task learning was used to predict both binary homologous recombination deficiency status and a continuous homologous recombination deficiency score. These choices make the unit of prediction and the use of multi-site pathology material central to interpreting the reported performance. (pmid:42907492)

On the reported test dataset, HRDPath using multi-task learning had an area under the receiver operating characteristic curve of 0.846, specificity of 0.938, and sensitivity of 0.692 for prediction of homologous recombination deficiency status. In the reported direct architecture comparison, DSMIL had an area under the receiver operating characteristic curve of 0.808, specificity of 0.875, and sensitivity of 0.538. Thus, within this test-dataset comparison, HRDPath had higher reported discrimination, specificity, and sensitivity than DSMIL. No confidence intervals or p-values for these performance estimates were located in the provided excerpts, so the precision of the reported differences cannot be assessed from this material. (pmid:42907492)

The benchmarking table also reports performance for several adapted architectures, including Vision Transformer, UNI, CLAM-MB, CLAM-SB, ZoomMIL, DeepHRD, and HRDPath with uni-task learning. The table indicates that the models were adapted with the study's patient-level and multiple-task learning approach for objective assessment. HRDPath's reported multi-task result therefore should be read as a result within a particular modeling and evaluation pipeline, not as a comparison with unmodified versions of every named architecture. The supplied excerpts identify the metrics but do not locate calibration measures, decision thresholds, or the denominator underlying the classification metrics. (pmid:42907492)

The investigators also conducted interpretability analyses. VAttentionCluster enrichment examined morphology and pathological features associated with homologous recombination deficiency-positive cancers and genomic alterations, and DBSCAN-based spatial clustering was used within the procedure. Anatomical-location comparisons examined tubo-ovarian, omental, other-pelvic-organ, and abdominal images. A paired-site analysis assessed retained whole-slide images from patients with evaluable profiles at both tubo-ovarian and omental sites. These analyses are descriptions of patterns identified by model attention and clustering in the analyzed pathology-image data; they do not independently establish the biological mechanism of homologous recombination deficiency. (pmid:42907492)

Analysis — Image-Based Classification Performance

The central observed pattern is that a model combining the reported architecture with patient-level aggregation and multi-task learning classified homologous recombination deficiency status more accurately by the listed test metrics than the DSMIL comparator. Independent validation on TCGAOV and PTRC strengthens the study design relative to evaluation solely in UWOV because it tests the model on datasets described as unseen during development. At the same time, the supplied results do not establish how performance varied by validation cohort, anatomical site, specimen characteristics, or the molecular reference method used to assign homologous recombination deficiency status. The analysis is therefore bounded to reported discrimination and classification performance in the study datasets. (pmid:42907492)

The study's selection process also matters. UWOV's multiple slides and multiple anatomical locations enabled patient-level aggregation and site-oriented analyses, whereas the independent cohorts are described as using diagnostic slides. That difference may affect the information available to the model across cohorts. The morphological associations from VAttentionCluster and DBSCAN-based analyses are consistent with attention-derived patterns in the retained pathology images under this design; they are not causal evidence that the identified features generate homologous recombination deficiency. Likewise, higher area under the receiver operating characteristic curve, specificity, and sensitivity than DSMIL are model-comparison results, not evidence of clinical utility or patient benefit. (pmid:42907492)

Limitations

The retrospective image-based development and evaluation design cannot establish causation. It evaluates associations between pathology-image features and assigned homologous recombination deficiency labels, rather than an intervention. Selection may arise from the composition of UWOV, the availability of diagnostic whole-slide images in TCGAOV and PTRC, and the decision to use UWOV for development while reserving the other cohorts for validation. Because UWOV included high-grade serous tubo-ovarian carcinoma and a small group of other ovarian cancers, while the validation cohorts are described as high-grade serous tubo-ovarian carcinoma, the studied populations are not fully identical. The provided excerpts do not locate comprehensive cohort-level label definitions or covariate distributions, limiting assessment of residual differences between development and validation data. (pmid:42907492)

Interpretability findings were also subject to analytic retention. The source states that some original UWOV whole-slide images were excluded from VAttentionCluster analyses because model attention did not form suitable spatially coherent clusters under the DBSCAN-based procedure. Consequently, the reported attention-cluster observations apply to the retained images meeting that procedure's requirements, not necessarily to every available slide. The source also lists a need for large training datasets as a limitation of HRDPath. Finally, the reported area under the receiver operating characteristic curve, sensitivity, and specificity do not alone establish clinical significance; the supplied material does not provide outcome evidence on care processes or patient outcomes. (pmid:42907492)