A deep learning approach for nucleus segmentation and tumor classification from lung histopathological images

Lung cancer is the leading cause of death worldwide. Early diagnosis is crucial to improve patients' chance of survival. Automated detection and analysis of cancer types can significantly improve the diagnosis process. It can aid treatment through follow-up analyses. This paper proposes a deep learning based pipeline for multi-class classification of lung tumor type (Benign (B), ADenoCarcinoma (ADC) and Squamous-Cell Carcinoma (SCC)) from histopathological images. A baseline classification method, the P-dir pipeline, is proposed where Whole Slide Histopathological Image (WSHI) patches are classified using the proposed Deep Convolutional Neural Network (DCNN) classifier. Since each cancer type is characterized by the difference in the structure of the nuceli, this research work proposes to improve the performance of classification by segmenting the nuclei. The P-seg pipeline is proposed to extract the nuclear regions from the WSHI patches using an Xception-style UNet based neural network, and this segmented region is then categorised into tumor types using the same downstream DCNN architecture. The classification system showed an accuracy of 95.40% and 99.60% using the Pdir and P-seg pipelines, respectively. The classification performed through Pseg pipeline exhibits significant improvement compared to the P-dir pipeline, supporting our hypothesis that nucleus segmentation improves classification performance. This paper posits that segmenting and retaining the nuclear regions in the input image to the tumor type classifier suppresses the importance of less relevant portions of the image during model training, pronounces nuclear region boundaries to highlight shape features, and reduces the overall computation cost of training. Ultimately, it induces a positive impact on classification performance. The enhanced performance obtained using the proposed Pseg pipeline supports our hypothesis.