Bayesian deep learning for reliable oral cancer image classification

Abstract

In medical imaging, deep learning-based solutions have achieved state-of-the-art performance. However, reliability restricts the integration of deep learning into practical medical workflows since conventional deep learning frameworks cannot quantitatively assess model uncertainty. In this work, we propose to address this shortcoming by utilizing a Bayesian deep network capable of estimating uncertainty to assess oral cancer image classification reliability. We evaluate the model using a large intraoral cheek mucosa image dataset captured using our customized device from high-risk population to show that meaningful uncertainty information can be produced. In addition, our experiments show improved accuracy by uncertainty-informed referral. The accuracy of retained data reaches roughly 90% when referring either 10% of all cases or referring cases whose uncertainty value is greater than 0.3. The performance can be further improved by referring more patients. The experiments show the model is capable of identifying difficult cases needing further inspection.

The Effectiveness of Artificial Intelligence in Detection of Oral Cancer

“…Seventeen studies were from various countries, with India accounting for 9 of them. 7 , 8 , 20 , 21 , 22 , 23 , 24 , 25 , 26 Three studies 27 , 28 , 29 were conducted in the United States, 2 30 , 31 were performed in China, and the other studies were carried out in Germany, 32 Taiwan, 33 and Jordan. 34 …”

“…All the selected studies were clinical trials. Nine were case-control studies, 7 , 8 , 20 , 21 , 23 , 28 , 30 , 33 , 34 7 were comparative studies, 22 , 24 , 25 , 26 , 27 , 29 , 31 and only one was a retrospective study, 30 with several of them employing various statistical procedures for a range of AI technologies.…”

“…In most investigations, deep learning has been used to detect OCs. Nonetheless, in terms of frequency of use, deep learning (convolutional neural network [CNN]) was used in 11 studies, 7 , 20 , 21 , 23 , 24 , 29 , 30 , 31 , 32 , 33 , 34 whilst 6 studies used machine learning. 8 , 22 , 25 , 26 , 27 , 28 The most frequently used subtype of the supervised machine learning approach is the support vector machine, which was used in 4 studies.…”

“…Uthoff et al sorted samples into the suspicious and nonsuspicious categories. 21 Other studies 8 , 20 , 23 , 24 , 27 , 28 , 29 , 31 offered an AI model that could categorise lesions as normal, precancerous, or cancerous, with or without additional categorisation of the samples into various stages of OC. Five studies 25 , 26 , 30 , 32 , 33 presented AI methods to categorise samples using binary classification as normal or malignant.…”

“…Also, they found that the immunopositive ratio values of both manual and automatic techniques were equal. 6 Song et al 23 India 2350 cheek mucosa images The intraoral data set of cheek mucosa images Learning machine: Bayesian deep network training = 300 epochs AC = 90% The performance can be further improved by referring more patients. The experiments show that the model is capable of identifying difficult cases needing further inspection.…”