Multi-Scale Deep Feature Fusion for Soft Tissue Sarcoma Subtype Classification From Histopathological Images

Abstract

Soft tissue sarcoma is a heterogeneous and a rare malignant tumour that is formed in connective tissues such as fat, muscle and fibrous tissue. For an efficient treatment planning accurate subtype classification is very essential. Traditional diagnostics rely on histopathological works which is time consuming and also subject to inter-observer variability. This research proposes a novel deep learning hybrid transformer architecture combined with EfficientNet V2B0, ConvNeXt V2(Tiny) and Light weight Swin Transformer for automated multi class STS subtype classification based on histopathological images which classifies 10 different sub types of Soft Tissue Sarcoma. Fine Grained textures are captured by EfficientnetV2B0, Local spatial features are extracted using the model ConvNeXt V2 and Global contextual dependencies are captured by using Swin Transformer. The proposed hybrid model demonstrates high precision of 0.91, accuracy 0.88, recall 0.88 and f1-score is of 0.88. High precision will always indicate that the model predicates very few false positive predictions. Effectiveness of deep learning methods helps in assisting pathologists and improves diagnostic reliability.