Developing Deep Learning Pipeline of Whole-Slide Images for Enhanced Diffuse Large B Cell Lymphoma (DLBCL) Subtyping and Outcome Prediction: Leveraging Self-Attention Transformer for Training and Inference

• Published in: Blood 2023-11, Vol.142 (Supplement 1), p.904-904
• Main Authors: Zhou, Chen, Xu, Jie, Prakash, Rishab, Torres-Cabala, Carlos P, Chen, Cheng-bang, Madduri, Kamesh, Rao, Arvind, Agasthya, Greeshma, Vega, Francisco, O'Malley, Dennis, Medeiros, L. Jeffrey, Kumara, Soundar, Iyer, Swami P.
• Format: Article
• Language: English
• ISSN: 0006-4971; 1528-0020
• DOI: 10.1182/blood-2023-187291

Background: Deep learning algorithms can help to analyze whole-slide images (WSI) in lymphoma pathology, identifying deeper features and patterns that may not be easily discernible to human observers. This pilot project is focused on diffuse large B-cell lymphoma (DLBCL), a heterogeneous disease with diverse genetic alterations. By leveraging self-attention trained clusters and transformers, the project aims to identify patterns and associations between mutational status and overall survival, potentially enhancing personalized treatment strategies and improving data reliability. Methods: This study employed a computer vision pipeline learning system to classify DLBCL subtypes using self-discovery of discriminatory features from scanned WSI. The workflow is shown in Figure 1a. First, we segmented tiles or patches from the gigapixel-sized WSI of 223 lymphoma cancer biopsy slides sourced from The Cancer Genome Atlas (TCGA), DLBCL, and Stanford DLBCL-Morph datasets. to optimize extraction of relevant features. For feature extraction, we utilized self-supervised pretraining with a Vision Transformer (ViT) network on a dataset of 1,515,000 patches. These patches were grouped into morphologically similar clusters using K-means, representing various lymphoma proliferation patterns. Dimensionality reduction with UMAP allowed for computational efficiency and feature visualization. Extracted features were used to predict overall survival using Bag-of-Words (BoWs). We further enhanced the model by incorporating geometric information, including nuclear characteristics from HoverNet. We assessed the model's performance using key metrics such as area under the curve (AUC) and accuracy. Additionally, genomic mutations with a frequency of 10% or higher from the TCGA were incorporated for enhanced predictive capability, including PIM1, SGK1, CARD11, KMT2D, SOCS1, BTG1, MUC16 and FAT4. Correlations and hierarchical clustering were performed on mutations, patches, and outcomes (Figure 1b). Results: Using the self-trained ViT encoder as the backbone and random forest, we demonstrate an accuracy of 0.88 on a weakly supervised task with all samples. The performance is significantly improved when compared to ResNet-50 trained on ImageNet. In addition, saliency maps from multi-attention heads provide excellent interpretations of morphological characteristics, including tumor stroma, cell location and necrosis. With the feature embeddings extracted by ViT, 10 morphologically distin