An AI-derived tumor cell density classification predicted which patients with locally advanced rectal cancer were more likely to benefit from irinotecan-intensified neoadjuvant chemoradiotherapy. Patients with higher tumor cell density demonstrated improved disease-free and overall survival outcomes with the intensified regimen vs those treated with standard chemoradiotherapy. Research findings from a post-hoc analysis of the phase III ARISTOTLE trial were published in eBioMedicine. 

“While the original trial showed little benefit from adding irinotecan, by using AI we found that we could distinguish patients who actually benefitted from those who did not. This demonstrates how AI can reveal tumor biology that is difficult to measure consistently by conventional means, and hopefully can be used to reveal additional insights that could lead to future treatments,” said lead author Zhuoyan Shen, PhD, a research fellow in the Department of Medical Physics and Biomedical Engineering at University College London (UCL). 

“This exciting AI study demonstrates the crucial importance of understanding the difference in biological behaviour of rectal cancer,” added co-author David Sebag-Montefiore, FRCP, FRCR, the Audrey and Stanley Burton Professor of Clinical Oncology at the University of Leeds.

Background and Study Methods 

Researchers developed a computational approach to quantifying tumor cell density within epithelial and stromal regions of tumors, and to assess if a response to treatment differs by tumor cell density status among patients with locally advanced rectal cancer who were receiving neoadjuvant chemoradiotherapy. The team developed a free online tool, called Octopath, where clinicians can upload biopsy slides to be analyzed.

In the post-hoc analysis, patients with available digitized pretreatment biopsies came from the multicenter, open-label, randomized controlled phase III ARISTOTLE trial, which evaluated adding irinotecan to chemoradiotherapy as preoperative treatment for patients with locally advanced rectal cancer. 

Tumor cell density was defined as the proportion of tumor cells in the tumor and stroma tissues. Tumor cell density was quantified with an AI framework that assessed digitized haematoxylin and eosin-stained whole-slide images. 

Patients were stratified by low or high tumor cell density with a cut-off value of 0.5 (50% of tumor cells). Tumor cell density status was combined with the treatment arms to stratify response and survival. 

Model Methods

The AI model was trained using large-scale open-source datasets and was tested on detailed microscopic images of 414 biopsy slides from patients with rectal cancer from the ARISTOTLE trial.

The model consisted of an EfficientNet-based tissue classifier and a pretrained SoftCTM cell detector, which was previously independently validated in a colorectal cancer cohort. Spatial data was augmented with random horizontal flipping and color augmentation. Tissue classification enabled whole-slide images to be segmented into nine colorectal tissue types and the cell detector determined tumor cell density.

Performance of the AI-based tumor cell density identification model was tested on 116 whole-slide images collected from a published study and on 100 resection samples from a cohort of the ARISTOTLE study that had already been annotated for an independent ongoing study. Annotations for the 116 whole-slide images were used as a reference standard for concordance assessments.

For stratification, the tumor cell density–highgroup was defined as patients with a tumor cell density of at least 0.5, and a low tumor cell density was below 0.5.

Key Findings 

The AI tissue classifier achieved a precision of 0.873 for tumor region identification and a recall of 0.984 for tumor-associated tissue identification. AI-determined tumor cell density was also highly concordant with manually assessed tumor cell density.

Forty-five percent of the participants in the post hoc analysis were considered tumor cell density–high.

A treatment–tumor cell density interaction was observed for disease-free survival (x² = 6.88; P = .009) and overall survival (x² = 10.61; P = .001), which was considered significant for both. 

Among patients with high tumor cell density, the addition of irinotecan to capecitabine and radiotherapy was associated a prolonged disease-free survival (hazard ratio [HR] = 0.57; 95% confidence interval [CI] = 0.36–0.90; P = .014), overall survival (HR = 0.50; 95% CI = 0.30–0.84; P = .008), and higher pathological complete response rates of 23% vs 11% (odds ratio [OR] = 2.46; 95% CI = 1.01–5.98; P = .042 [P = .13 after adjustment for multiple testing]) vs standard capecitabine and radiotherapy. 

In the low tumor cell density group, on the other hand, no significant difference was observed in disease-free survival between the treatment arms (HR = 1.29; 95% CI = 0.85–1.95; P = .22). Additionally, patients in the investigational treatment arm who had low tumor cell density showed a trend toward worse overall survival (HR = 1.55; 95% CI = 0.96–2.51; P = .07). No significant difference was noted in pathologic complete response rates either (14% vs 22%; OR = 0.60; 95% CI = 0.28–1.29; P = .19). 

“Intensifying already taxing treatments puts additional strain on patients suffering from cancer. Clinicians need reliable ways to identify who is most likely to benefit before such treatments begin so potential side effects are avoided. Our findings show that doctors assisted by AI can pinpoint which patients will likely benefit from the more intensive treatment before it begins,” said senior author Maria A. Hawkins, MD, MCRP, FRCR, a consultant clinical oncologist and Professor of Radiation Oncology, Department of Medical Physics and Biomedical Engineering, UCL. 

DISCLOSURES: Cancer Research UK Radiation Research Network - Project Seed Funding, Cancer Research UK ARISTOTLE sample collection grant, UK Research and Innovation Future Leadership Fellowship and the Radiation Research Unit at the Cancer Research UK City of London Centre Award. For full disclosures of the study authors as well as code availability for the AI framework, visit thelancet.com. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.