Unsupervised Deep-Learning on H&E Images of Primary Melanoma Tumors for BRAF Mutational Status and Survival Prediction: An Enrichment for Clinical Decision Making

01 MAR 2023

Abstract

Background and aims

We designed an un-supervised deep neural network that extracts features from no-padding patches of WSI (tiles of size 224x224 pixels) using a self-supervised learning framework, which are then fed into a multiple classifier assigning a BRAF mutational status probability to each WSI (1 patient = 1 WSI). The model was trained and validated on 223 cases from the TCGA dataset. Performance was tested on a left-out TCGA set (n=54) as well as an external set of 49 cases from patients of our institution (patients with available cutaneous primary melanoma tissue, Braf mutational status and 5 years OS, included from April 2014 to April 2016). The slides were digitized using a whole slide scanner (Roche Ventana iSan HT). Any manual annotation was performed on WSI. We carry out an investigation of the characterisation of optimal separative cut-offs for the classification of the mutational status outputs with maximal positive (PPV) and negative (NPV) predictive values. The same model is then adapted to handle the prediction of survival risk scores of the patients.

Patients and Methods

Of the 223 TCGA cases, 120 were BRAF-mutated (BM, 53.8%) and 103 BRAF wild-type (BW, 46.2%) and the external dataset comprises 19 BM (38.8%) and 30 cases with BW (61.2%). The prediction of the BRAF mutational status on the basis of the morphological features achieves an AUC of 0.69 and 0.67 respectively on the TCGA and external test datasets. For the latter (n=49) and based on an adequate selection of the cut-off for classification decision, the proposed model succeeds in predicting the mutational status for 44.90% of the cases with a NPV and PPV of 100%. Concerning prediction of survival risk scores, the proposed pipeline yielded c-indices of 0.65 and 0.62 respectively on the TCGA and external dataset, which are comparable to the reported scores for the other types of tumors based on deep learning frameworks.

Results

This pilot study in melanoma clearly shows that this infrequent unsupervised deep-learning approach to HE image analysis is able to discover new digital predictive and prognostic biomarkers. It has the advantage of leaving the region of interest selection entirely to the deep network. By this way, it is not biased by the pathologist annotation for tumor region and can take into account the background and the interface with the non tumor tissue. Thus, it could have in the near future the potential to accelerate and improve the clinical decision-making process in the field of melanoma. Also, these findings will undoubtedly open room for substantial applications with economic and treatment delay fallouts.

Discussion / Conclusion

However, as the sample size is small, training on larger and independent datasets is needed to boost the performance of our method, especially to better prescreen patients for BRAF mutational status before genetic testing. Our work provide additional insights on how machine learning-based WSI analysis has the potential to complete and improve the predictive and prognostic approaches for the pathologist.

Read the full paper

Try it on your cases ↗︎