Grading Tumor Malignancy via Deep Bidirectional LSTM on Graph Manifold Encoded Histopathological Image

摘要

Shortage of clinicians in developing countries demands computer aided histopathological image (HI) classification systems for breast cancer (BC) taxonomy. These expert dependent diagnosis are cumbersome & subjective in nature. As such, the importance of efficient feature engineering to differentiate non-uniform color distribution and wide texture variability of cancerous tissues cannot be overemphasized. Recently, computationally intensive deep learning methods have shown efficacy in overcoming these challenges. In this work, an automated BC categorization framework employing Bidirectional LSTM (Bi-LSTM) on manifold encoded HI of the quasi-isometric topological space is proposed. Manifold learning via Landmark ISOMAP (L-ISOMAP) on stain normalized unfolded HI's have been used to model its latent multidimensional structural dynamics. Thereafter, these manifold embedded HI point-clouds are fed to a deep Bi-LSTM recurrent neural network (RNN) to comprehend its complex spatial dependencies and apprehend the underlying local and global contextual morphology. Our proposed method has been validated on BreakHis, a large-scale publicly available dataset comprising of 7,909 HI of both benign and malignant classes. Further, manifold embedding manifests in lower computational complexity for the deep learning phase, while the average classification rate of 97.2% outperforms the existing state-of-the-art and validate its diagnostics adequacy for clinical settings deployment.