Enhancing the Stability and Quality Assessment of Visual Explanations for Thorax Disease Classification Using Deep Learning

Date

2023-9

Author

Rahimiaghdam, Shakiba

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

205
views

91
downloads

Cite This

Deep learning models are renowned but their complex internal workings often render them opaque. While efforts are underway to amplify their explainability, a substantial gap persists. One pivotal challenge is the unavailability of quantifiable metrics for evaluating visual explanations, leading to a reliance on manual assessments or suboptimal metrics. This restricts scalability, compromises reproducibility, and undermines trustworthiness. The necessity for objective metrics is further underscored during hyperparameters fine-tuning and the evaluation of various Explainable Artificial Intelligence (XAI) models. Another prominent hurdle is the instability exhibited by models like Local Interpretable Model-agnostic Explanations (LIME). The random perturbations intrinsic to LIME lead to inconsistent explanations, eroding trust and obstructing their integration into critical applications. To navigate these challenges, we unveil a robust method for the objective assessment, refinement, and juxtaposition of visual explanation algorithms, offering a concrete solution to metric inadequacy. Addressing instability, we introduce MindfulLIME. This cutting-edge approach leverages graph-based pruning and uncertainty sampling, strategically crafting purposeful samples and elevating the reliability and consistency of visual explanations - a discernible advancement over LIME. Our intricate qualitative analysis, applied to unseen random samples from the VinDr-CXR dataset, attests to the preeminence of our selected metric. In a comparative analysis involving multi-label, multi-class diagnosis of thorax diseases, MindfulLIME, evaluated via this refined metric, epitomizes optimal stability and augmented localization accuracy. This underscores its capacity to elevate the trust quotient associated with machine learning applications in the nuanced field of medical imagery.

Subject Keywords

Explainable Artificial Intelligence (XAI), Visual Explanation Evaluation, Machine Learning, Deep Learning Models, Thorax Diseases Diagnosis

URI

https://hdl.handle.net/11511/105447

Collections

Graduate School of Natural and Applied Sciences, Thesis