by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a recent study, researchers at Wesleyan University (Connecticut, USA) and  Gachon University Gil Hospital (Incheon, Republic of Korea) developed a deep learning framework to predict mismatch repair deficiency (MMR-D) status using digitized pathology slides from patients with endometrial cancer.

The findings suggest that the proposed deep learning model can predict an important molecular biomarker from whole slide images (WSIs) of routine pathology slides, which could enable precision oncology if validated at larger scales.

“Our results suggest that this multi-resolution ensemble learning approach can swiftly predict MMR status from pathological WSIs, aiding in the quick determination of suitable treatment plans, particularly the use of immunotherapy for patients with MMR-D endometrial cancer,”

said Jisup Kim, MD, assistant professor at Gachon University College of Medicine and the corresponding author of the study.

The report was published in the Journal of Imaging Informatics in Medicine.

Rationale: Automating MMR-D Detection

Endometrial cancer, a malignancy affecting the lining of the uterus, is a heterogeneous disease with varying molecular subtypes and treatment responses. One subtype, characterized by MMR-D, has been shown to respond favorably to immunotherapy. This is because MMR-D leads to microsatellite instability (MSI), making tumors more immunogenic and, therefore, amenable to immunotherapy.

However, current diagnostic methods for MMR-D, primarily based on immunohistochemistry (IHC) and molecular testing, are time consuming and resource intensive, often delaying the initiation of therapy.

A team of researchers developed an automated solution for MMR-D detection using deep learning models trained on digitized pathology slides of tissues stained with hematoxylin and eosin (H&E). By harnessing the power of AI, the researchers aimed to create a diagnostic tool that could streamline the identification of patients eligible for immunotherapy, thereby reducing diagnostic delays and improving treatment outcomes.

“The study was driven by the increasing routine process of testing for MMR status in endometrial carcinoma for molecular classification. MMR-D, which affects over 30% of endometrial cancer patients, can significantly influence treatment decisions, notably in favoring immunotherapy for individuals with MMR-D,”

explained Dr. Kim.

Approach: Harnessing the Power of Multi-Resolution Ensemble Learning

The researchers assembled a dataset of 1,168 WSIs from 325 patients with endometrial cancer collected at Gachon University Gil Medical Center. Images were manually labeled by pathologists as positive or negative for MMR-D based on the immunohistochemistry results for four key MMR proteins (MLH1, MSH2, MSH6, and PMS2).

To account for variations in H&E staining intensity, the team employed a CycleGAN-based network to normalize the color distribution across the WSI tiles extracted at multiple magnifications (2.5´, 5´, and 10´). This multi-resolution approach enabled the model to capture both broad structural patterns and fine cellular details, thereby enhancing its predictive accuracy.

Three distinct deep learning architectures, InceptionResNetV2, EfficientNetB2, and EfficientNetB3, were trained on the normalized tiles, each capturing features from a different magnification level. The ensemble model aggregated these multi-scale representations by leveraging the strengths of each architecture to enhance predictive power.

“The most innovative aspect of the study was the application of multi-resolution ensemble learning in digital pathology to accurately predict MMR status. We used three different networks for each magnification level and integrated these through ensemble learning to predict the MMR status. Leveraging information from different magnification levels of WSIs, this approach allowed for a comprehensive analysis incorporating features across multiple resolutions to accurately predict MMR status,”

Dr. Kim explained.

Proof-of-Concept Validation

When evaluated on a hold-out test set of 234 WSIs, the overall area under the receiver operating characteristic curve (AUC) range across the three models was 0.767 to 0.821, indicating good performance in predicting MMR-D at the whole slide level.

The ensemble model powered by the EfficientNetB2 architecture performed the best among the three models, providing an AUC of 0.821 (95% confidence interval [CI]: 0.763–0.879) for predicting MMR-D status at the whole slide level. Moreover, the model provided an accuracy of 0.778 (95% CI: 0.718–0.838), sensitivity of 0.827 (95% CI: 0.769–0.885), and specificity of 0.764 (95% CI: 0.712–0.816).

“This high AUC indicates the model’s strong potential in distinguishing between MMR-D and MMR-proficient cases and correctly identifying patients who might benefit from immunotherapy,”

noted Dr. Kim.

Future Work

“While the study achieved significant results, it also highlighted areas for future research. Questions remain about the model’s generalizability across datasets from different institutions and its interpretability in clinical settings.Future studies will focus on validating the model with external datasets and further refining the approach for enhanced clinical applicability and understanding,”

Dr. Kim acknowledged.

In addition, the authors proposed incorporating human-interpretable feature extraction techniques to enhance the model’s transparency and robustness in clinical settings.

Looking ahead, the researchers envision extending their approach to other molecular subtypes of endometrial cancer and exploring its applicability to different cancer types. Ultimately, their goal is to contribute to the realization of precision oncology, in which tailored treatments are guided by accurate and automated biomarker detection, empowering clinicians to provide the right therapy for the right patient at the right time.

According to Dr. Kim, by automating the detection of MMR-D status directly from routine pathology slides, clinicians could rapidly identify patients with endometrial cancer who are candidates for immunotherapy without the need for additional molecular testing. This streamlined process could accelerate treatment decision making, potentially improving patient outcomes.

The study was supported by the Technology Innovation Program funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea), the Korea Medical Device Development Fund grant funded by the Korean government, and the Gachon University Research Fund.

References

1. Whangbo J, Lee YS, Kim YJ, Kim J, Kim KG. Predicting Mismatch Repair Deficiency Status in Endometrial Cancer through Multi-Resolution Ensemble Learning in Digital Pathology. J Imaging Inform Med. Published online February 20, 2024. doi:10.1007/s10278-024-00997-z