ResNet for Histopathologic Cancer Detection, the Deeper, the Better?

Abstract

Histopathological image classification has become one of the most challenging tasks among researchers due to the categories and fine-grained variability of the disease. Based on deep residual convolutional neural networks (ResNet), we study the classification tasks in medical image datasets. In this paper, we seek to answer the following central question in the context of medical image analysis: Is the deeper the network layer, the better the performance? Can the transfer of pre-trained deep ResNet with sufficient fine-tuning of all layers be better than the transfer with freezing most layers and training only the last layer? To address this question, ResNet with 18, 34, 50, and 152 layers pretrained from ImageNet were transferred in two different strategies on the histopathology image datasets of Kaggle, which included 220,025 histopathology patches with labels. The performance was also compared with traditional machine learning models. Our experiments consistently demonstrated that: (1) the deeper the ResNet network layers are, the better performance may not be (area under curve (AUC) value of ResNet-34 is 0.992 and ResNet-152 is 0.989). The performance should be determined by the richness of semantic features of the datasets but not the depth of the layers and the time and resources required for training should also be considered in model selection. (2) Transfer learning with deep ResNet (AUC: 0.881~0.992) works better than traditional machine learning with logistic regression (AUC: 0.775), which indicates the better generalization ability and robustness of the deep ResNet. (4) Freezing most layers of the deep ResNet and only train the last fully connected layer does not improve the accuracy and efficiency of transfer learning. Fine-tuning all-layer scheme offers a practical way to reach the best performance based on the amount of available data. (5) The performance of both transfer strategies depends largely on datasets. In conclusion, both strategies produce good results in terms of training time and overall accuracy when compared to models trained from scratch or traditional machine learning models.

Received: 17 January 2023 | Revised: 27 February 2023 | Accepted: 28 February 2023

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support the findings of this study are openly available in [Kaggle HCD] at https://www.kaggle.com/datasets/drbeane/hcd-cropped