A clinical decision support system optimising adjuvant chemotherapy for colorectal cancers by integrating deep learning and pathological staging markers: a development and validation study.

BACKGROUND: The DoMore-v1-CRC marker was recently developed using deep learning and conventional haematoxylin and eosin-stained tissue sections, and was observed to outperform established molecular and morphological markers of patient outcome after primary colorectal cancer resection. The aim of the present study was to develop a clinical decision support system based on DoMore-v1-CRC and pathological staging markers to facilitate individualised selection of adjuvant treatment. METHODS: We estimated cancer-specific survival in subgroups formed by pathological tumour stage (pT<4 or pT4), pathological nodal stage (pN0, pN1, or pN2), number of lymph nodes sampled (≤12 or >12) if not pN2, and DoMore-v1-CRC classification (good, uncertain, or poor prognosis) in 997 patients with stage II or III colorectal cancer considered to have no residual tumour (R0) from two community-based cohorts in Norway and the UK, and used these data to define three risk groups. An external cohort of 1075 patients with stage II or III R0 colorectal cancer from the QUASAR 2 trial was used for validation; these patients were treated with single-agent capecitabine. The proposed risk stratification system was evaluated using Cox regression analysis. We similarly evaluated a risk stratification system intended to reflect current guidelines and clinical practice. The primary outcome was cancer-specific survival. FINDINGS: The new risk stratification system provided a hazard ratio of 10·71 (95% CI 6·39-17·93; p<0·0001) for high-risk versus low-risk patients and 3·06 (1·73-5·42; p=0·0001) for intermediate versus low risk in the primary analysis of the validation cohort. Estimated 3-year cancer-specific survival was 97·2% (95% CI 95·1-98·4; n=445 [41%]) for the low-risk group, 94·8% (91·7-96·7; n=339 [32%]) for the intermediate-risk group, and 77·6% (72·1-82·1; n=291 [27%]) for the high-risk group. The guideline-based risk grouping was observed to be less prognostic and informative (the low-risk group comprised only 142 [13%] of the 1075 patients). INTERPRETATION: Integrating DoMore-v1-CRC and pathological staging markers provided a clinical decision support system that risk stratifies more accurately than its constituent elements, and identifies substantially more patients with stage II and III colorectal cancer with similarly good prognosis as the low-risk group in current guidelines. Avoiding adjuvant chemotherapy in these patients might be safe, and could reduce morbidity, mortality, and treatment costs. FUNDING: The Research Council of Norway.

Deep learning for prediction of colorectal cancer outcome: a discovery and validation study.

BACKGROUND: Improved markers of prognosis are needed to stratify patients with early-stage colorectal cancer to refine selection of adjuvant therapy. The aim of the present study was to develop a biomarker of patient outcome after primary colorectal cancer resection by directly analysing scanned conventional haematoxylin and eosin stained sections using deep learning. METHODS: More than 12 000 000 image tiles from patients with a distinctly good or poor disease outcome from four cohorts were used to train a total of ten convolutional neural networks, purpose-built for classifying supersized heterogeneous images. A prognostic biomarker integrating the ten networks was determined using patients with a non-distinct outcome. The marker was tested on 920 patients with slides prepared in the UK, and then independently validated according to a predefined protocol in 1122 patients treated with single-agent capecitabine using slides prepared in Norway. All cohorts included only patients with resectable tumours, and a formalin-fixed, paraffin-embedded tumour tissue block available for analysis. The primary outcome was cancer-specific survival. FINDINGS: 828 patients from four cohorts had a distinct outcome and were used as a training cohort to obtain clear ground truth. 1645 patients had a non-distinct outcome and were used for tuning. The biomarker provided a hazard ratio for poor versus good prognosis of 3·84 (95% CI 2·72-5·43; p<0·0001) in the primary analysis of the validation cohort, and 3·04 (2·07-4·47; p<0·0001) after adjusting for established prognostic markers significant in univariable analyses of the same cohort, which were pN stage, pT stage, lymphatic invasion, and venous vascular invasion. INTERPRETATION: A clinically useful prognostic marker was developed using deep learning allied to digital scanning of conventional haematoxylin and eosin stained tumour tissue sections. The assay has been extensively evaluated in large, independent patient populations, correlates with and outperforms established molecular and morphological prognostic markers, and gives consistent results across tumour and nodal stage. The biomarker stratified stage II and III patients into sufficiently distinct prognostic groups that potentially could be used to guide selection of adjuvant treatment by avoiding therapy in very low risk groups and identifying patients who would benefit from more intensive treatment regimes. FUNDING: The Research Council of Norway.

Designing deep learning studies in cancer diagnostics.

The number of publications on deep learning for cancer diagnostics is rapidly increasing, and systems are frequently claimed to perform comparable with or better than clinicians. However, few systems have yet demonstrated real-world medical utility. In this Perspective, we discuss reasons for the moderate progress and describe remedies designed to facilitate transition to the clinic. Recent, presumably influential, deep learning studies in cancer diagnostics, of which the vast majority used images as input to the system, are evaluated to reveal the status of the field. By manipulating real data, we then exemplify that much and varied training data facilitate the generalizability of neural networks and thus the ability to use them clinically. To reduce the risk of biased performance estimation of deep learning systems, we advocate evaluation in external cohorts and strongly advise that the planned analyses, including a predefined primary analysis, are described in a protocol preferentially stored in an online repository. Recommended protocol items should be established for the field, and we present our suggestions.