A 3 Tier CNN model with deep discriminative feature extraction for discovering malignant growth in multi-scale histopathology images. (2021)

• Record Type: Journal Article
• Authors: Kate, Vandana; Shukla, Pragya
• Is Part Of: Informatics in medicine unlocked. Volume 24(2021)
• Publication Date: 2021
• Subjects: Image-multi-classification -- Histopathology-images -- Breast- cancer (BC) -- Feature-extraction -- Deep-neural-networks; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2021.100616
• Languages: English
• ISSNs: 2352-9148

Abstract

Abstract: The Convolutional Neural Network (CNN) is intended to generalize and automatically learn spatial hierarchies of features, using stacked convolution-pooling layers with Backpropagation of errors. Two predominant challenges confronted while applying the CNN model for cancer classification tasks are limited size of dataset and model overfitting. The proposed 3 Tier CNN model employs- (i) spatial attention and attention across channels that potentially improves input representation power of the network, (ii) enforces Separable Convolutions for training that alleviates data overfitting concerns and shrinks network complexity, (iii) uses skip connections in some sub-networks to boost gradient flow during backpropagation and (iv) the model additionally combines spatial and semantic information present in early, intermediate and final layers which are passed to dense layers for producing better classification results. The results of the proposed model when applied to Breast Cancer histopathological imaging modality yielded the best overall performance with 98.5% average accuracy for 200X scaled images. When various images of different scales (40X/100X/200X/400X) were integrated to deliberately expand the input dataset making it appropriate for multiclass classification, the model gave an average accuracy of 95.5% and 92.8% for benign and malignant subclass classification. As the model considerably reduces the domain know-how of a specific problem, it can be used to attain better classification results for similar problems.