Double-Balanced Loss for Imbalanced Colorectal Lesion Classification.
Chang Yu1, Wei Sun1, Qilin Xiong1
1Information Engineering College, Shanghai Maritime University, Shanghai 201306, China.
Computational and Mathematical Methods in Medicine
|August 18, 2022
Summary
This study introduces a novel double-balanced loss function to improve deep learning models for colorectal cancer detection from colonoscopy images. This method addresses data imbalance, enhancing diagnostic accuracy for better patient outcomes.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Colorectal cancer (CRC) is a global health concern with improved survival rates linked to early detection.
- Deep learning-based computer-aided diagnosis (CADx) systems show promise in analyzing colonoscopy images to reduce diagnostic errors.
- A key challenge in medical image analysis is the inherent class imbalance in training datasets, which can compromise deep learning model performance.
Purpose of the Study:
- To develop and evaluate a novel loss function designed to mitigate the impact of data imbalance in deep learning models for colorectal cancer detection.
- To address both sample size and sample difficulty imbalances within medical image datasets.
- To enhance the accuracy and reliability of computer-aided diagnosis systems for colorectal cancer screening.
Main Methods:
- A new loss function, termed the "double-balanced loss function," was proposed for deep learning models.
- This loss function incorporates considerations for sample size and sample difficulty into the loss calculation.
- The proposed method was integrated into a deep learning framework for the medical diagnosis of colorectal cancer using colonoscopy images.
Main Results:
- The double-balanced loss function demonstrated superior performance in handling imbalanced classification tasks involving colorectal medical images.
- Experimental validation using three distinct colorectal white-light endoscopic image datasets confirmed the effectiveness of the proposed approach.
- The novel loss function improved the impact of datasets on classification accuracy, outperforming existing methods on imbalanced data.
Conclusions:
- The proposed double-balanced loss function is an effective strategy for improving deep learning model performance on imbalanced colorectal cancer image datasets.
- This approach offers a valuable tool for enhancing the accuracy of computer-aided diagnosis in gastroenterology.
- Further development and application of balanced loss functions can significantly advance automated medical image analysis and diagnostics.
