CS-AF: A Cost-sensitive Multi-classifier Active Fusion Framework for Skin Lesion Classification (2004.12064v2)

Published 25 Apr 2020 in cs.CV, cs.LG, and eess.IV

Abstract: Convolutional neural networks (CNNs) have achieved the state-of-the-art performance in skin lesion analysis. Compared with single CNN classifier, combining the results of multiple classifiers via fusion approaches shows to be more effective and robust. Since the skin lesion datasets are usually limited and statistically biased, while designing an effective fusion approach, it is important to consider not only the performance of each classifier on the training/validation dataset, but also the relative discriminative power (e.g., confidence) of each classifier regarding an individual sample in the testing phase, which calls for an active fusion approach. Furthermore, in skin lesion analysis, the data of certain classes (e.g., the benign lesions) is usually abundant making them an over-represented majority, while the data of some other classes (e.g., the cancerous lesions) is deficient, making them an underrepresented minority. It is more crucial to precisely identify the samples from an underrepresented (i.e., in terms of the amount of data) but more important minority class (e.g., certain cancerous lesion). In other words, misclassifying a more severe lesion to a benign or less severe lesion should have relative more cost (e.g., money, time and even lives). To address such challenges, we present CS-AF, a cost-sensitive multi-classifier active fusion framework for skin lesion classification. In the experimental evaluation, we prepared 96 base classifiers (of 12 CNN architectures) on the ISIC research datasets. Our experimental results show that our framework consistently outperforms the static fusion competitors.

Authors (3)

Di Zhuang (14 papers)
Keyu Chen (76 papers)
J. Morris Chang (23 papers)

Citations (15)

View on Semantic Scholar

Summary

CS-AF: Cost-Sensitive Multi-classifier Active Fusion for Skin Lesion Classification

Introduction

In the domain of skin lesion analysis, the superiority of Convolutional Neural Networks (CNNs) in achieving state-of-the-art performance is well-established. However, employing a single CNN classifier approach has its limitations, especially when dealing with limited and statistically biased datasets typical of this field. The fusion of results from multiple classifiers emerges as a more potent and robust alternative. This paper introduces a novel cost-sensitive multi-classifier active fusion framework named CS-AF, specifically designed to tackle the classification of skin lesions. The paper meticulously highlights CS-AF's considerations of individual classifier performance, confidence levels on unseen samples, and importantly, class imbalance with special attention to underrepresented but critical minority classes.

Methodological Approach

The framework is predicated on two principal components: objective and subjective weights, which collectively inform the fusion process.

Objective weights are derived from the classifiers' reliability based on their performance metrics on a validation set. Moreover, these weights incorporate a cost-sensitive mechanism through a customizable cost matrix which factors in the costs associated with misclassification, particularly emphasizing the severe repercussions of incorrectly classifying critical skin conditions.
Subjective weights are calculated based on a classifier's confidence in predicting classes of individual unseen samples during the testing phase, enabling dynamic weight adjustments that cater to sample-specific scenarios.

The architecture involves training 96 base classifiers across 12 different CNN designs on the International Skin Imaging Collaboration (ISIC) datasets. These datasets primarily feature images of skin lesions, with an emphasis on a balanced representation across various skin conditions.

Experimental Evaluation

The CS-AF framework was subjected to rigorous testing against static fusion counterparts and a version of active fusion devoid of the cost-sensitive feature. The evaluation showcased CS-AF's consistent superiority in accuracy and effectiveness in minimizing total cost - a testament to its intricate design that intelligently balances accuracy with the crucial aspect of cost sensitivity.

For varying counts of classifiers fused, CS-AF demonstrated an improvement of 2%-5% in accuracy over the highest-performing individual CNN classifier.
Furthermore, the framework's adaptability to different cost matrices was proven, with varying emphasis either on cancerous or benign lesions, CS-AF achieved the lowest total costs, confirming its capacity to adjust based on specified cost sensitivities.

Implications and Speculations on Future Developments

The implications of CS-AF in the field of skin lesion analysis are noteworthy, with potential applicability extending into other medical imaging disciplines. The framework not only advances the accuracy of classifications but does so with an acute awareness of the varying severities of misclassifications, mirroring real-world medical diagnostic scenarios where the cost of error can indeed be high.

One might speculate that future iterations of CS-AF could explore alternative metrics to further refine objective weights. Also, integrating a learning-based mechanism for ascertaining subjective weights could enhance the framework's adaptability and precision. Expansion of this framework's application to other medical imaging tasks could also illuminate its broader utility and impact, potentially revolutionizing diagnostic procedures by harnessing the full capabilities of AI.

Conclusion

CS-AF emerges as a pioneering framework in skin lesion classification, meticulously designed to enhance accuracy while concurrently addressing the vital concerns of cost sensitivity. This framework sets a new benchmark in the application of multi-classifier fusion techniques, particularly in medical imaging domains characterized by class imbalances and the critical need for precision. Through this innovative amalgamation of objective and subjective elements, informed significantly by cost sensitivities, CS-AF paves the way for future research and application paradigms in AI-driven diagnostics.

PDF Markdown

Related Papers

Find Related Papers