A Review of Deep Neural Networkbased Uncertainty Quantification Methods for the Classification of Breast Cancer

EasyChair Preprint 9686

Abstract

In recent years, deep learning-based technologies have become widely used in the medical area with
remarkable success. The output of many of these methods, however, has excessive confidence levels
and the majority of them cannot provide numerical guarantees. There is no way they could be
effective, and they might even cause permanent harm. Therefore, the approximation of Bayesian
and Ensemble learning techniques are considered as uncertainty quantification approaches to take
on such a problem. In this study, we implement and assess three UQ models for categorising breast
tumour tissues. A few examples of these techniques include the Bayesian Ensemble, the MCD
Ensemble, and the Mont Carlo Dropout (MCD) approach. In addition, the present study takes into
account a transfer learning technique and a pre-trained CNN in order to boost the classification's
accuracy and remove the negative effects of the study's small data collection in Wisconsin Diagnostic
Breast Cancer (WDBC). Novel performance criteria are used to assess estimated uncertainty, and the
three proposed models are compared based on their capacity to quantify the reliability of
classification. In the study, we conducted quantitative and qualitative analyses to indicate that
models exhibit substantial ambiguity in misclassifications, which is critical for establishing the
frequency of medical diagnosis hazards. Therefore, we hope to determine whether the deep neural network's output can be trusted by applying these new evaluation criteria. Further, the Bayesian
Ensemble model's uncertainty quantification is shown to be more trustworthy through the analysis.

Keyphrases: Bayesian ensemble, CNN, Medical Diagnosis, Mont Carlo Dropout, Wisconsin Diagnostic Breast Cancer

@booklet{EasyChair:9686,
  author    = {Agha Salman Haider and Alighazi Siddiqui and Md Imran Alam and Felipe de Castro Dantas Sales and Shams Tabrez Siddiqui and V Vijayabhaskar and Ranju Lal and Harpreet Kaur},
  title     = {A Review of Deep Neural Networkbased Uncertainty Quantification Methods for the Classification of Breast Cancer},
  howpublished = {EasyChair Preprint 9686},
  year      = {EasyChair, 2023}}