Fine-Tuning and Training of DenseNet for Histopathology Image Representation Using TCGA Diagnostic Slides

Abstract

Feature vectors provided by pre-trained deep artificial neural networks have become a dominant source for image representation in recent literature. Their contribution to the performance of image analysis can be improved through finetuning. As an ultimate solution, one might even train a deep network from scratch with the domain-relevant images, a highly desirable option which is generally impeded in pathology by lack of labeled images and the computational expense. In this study, we propose a new network, namely KimiaNet, that employs the topology of the DenseNet with four dense blocks, fine-tuned and trained with histopathology images in different configurations. We used more than 240,000 image patches with 1000×1000 pixels acquired at 20× magnification through our proposed "highcellularity mosaic" approach to enable the usage of weak labels of 7,126 whole slide images of formalin-fixed paraffin-embedded human pathology samples publicly available through the The Cancer Genome Atlas (TCGA) repository. We tested KimiaNet using three public datasets, namely TCGA, endometrial cancer images, and colorectal cancer images by evaluating the performance of search and classification when corresponding features of different networks are used for image representation. As well, we designed and trained multiple convolutional batch-normalized ReLU (CBR) networks. The results show that KimiaNet provides superior results compared to the original DenseNet and smaller CBR networks when used as feature extractor to represent histopathology images.

Pay Attention with Focus: A Novel Learning Scheme for Classification of Whole Slide Images

“…We fined-tuned the DenseNet on training patches using weak labels obtained from their respective WSIs. The weakly labelled fine-tuning has shown to be effective [19]. In our case, the weak labels are anatomic site, and primary diagnosis, arranged in a hierarchy.…”

“…The vertical classification results are reported in Table 2 4 . The results show that FocAtt-MIL can elevate the accuracy of pre-trained fea- tures; DenseNet features have shown to under-perform compared to KimiaNet features [19,12]. However, within the proposed FocAtt-MIL scheme, DenseNet features become quite competitive.…”

“…Pan-cancer Analysis -In the second experiment series, we evaluated the approach against a large-scale pan-cancer classification of WSIs. The dataset used for this task has been proposed by Riasatian et al [19]. It comprises more than 7 TB data, consisting of 7,097 training, and 744 test WSIs, distributed across 24 different anatomic sites, and 30 different primary diagnoses.…”

“…All WSIs in the dataset are taken from a public repository of WSIs, TCGA [23]. We obtained a mosaic for each WSI, and then applied a cellularity filter [19] to further reduce the number of patches in each mosaic. Subsequently, we obtained 242,202 patches for training WSIs and 116,088 patches for testing WSIs.…”

“…We prepared a separate "bag" for each feature extractor. These three feature extractors are: DenseNet (DN) [9], KimiaNet [19], and the fine-tuned DenseNet (FDN). We fined-tuned the DenseNet on training patches using weak labels obtained from their respective WSIs.…”