Medicine XAI May - August 2022

1. The article states that recognizing lung sound patterns is critical to detecting and monitoring respiratory disease. Current techniques for breath sound analysis used by industry experts are subject to interpretation. Therefore, a precise and automatic breath sound classification system is desirable. In this work, we took a data-driven approach to classify abnormal lung sounds. The authors compared the performance with three different functions extraction techniques, which are Short Time Fourier Transform (STFT), Mel and Wav2vec spectrograms, as well as three different classifiers including pre-trained ResNet18, LightCNN, and audio spectrogram transformer. The authors' fundamental contribution contains a comparative analysis of different audio feature extractors and classifiers based on neural networks and their implementation of a complete pipeline with STFT and matched ResNet18 Network. The method proposed has achieved Harmonic Scores at the levels of 0.89, 0.80, 0.71, 0.36 for tasks and 1-1, 1-2, 2-1 and 2-2 on the 2022 IEEE BioCAS Grand Challenge Test Sets on the classification of breath sounds.

2. The aim of the study was to use high-resolution quantitative CT (QCT) imaging to predict the diagnosis and prognosis of fibrosing interstitial lung disease (ILD).  The authors' approach: 40 patients with ILD (20 typical interstitial pneumonitis (UIP), 20 ILD without UIP) were classified by expert consensus by 2 radiologists and followed up for 7 years during which clinical variables were recorded. After lung field segmentation, a total of 26 texture features were extracted using a network-based approach (TM model). The TM model was compared to the previous one based on histograms (HM) for their ability to rank UIP versus non-UIP. For prognostic assessment, a survival analysis was performed by comparing specialized diagnostic labels with TM measurements. The authors' results were as follows: In the classification analysis, the TM model was superior to the HM method with an AUC of 0.70. While the survival curves for the UIP and non-UIP expert labels were not statistically different in the Cox regression analysis, the TM-QCT features allowed a statistically significant cohort assignment. The authors' conclusions were that the TM model was superior to the HM model in distinguishing between UIP and non-UIP models. More importantly, since MT allows for the subdivision of the cohort into different survival groups, the UIP vs non-UIP QCT TM models can improve the diagnosis of interstitial lung disease and provide more accurate prognosis and better patient management.

3. This paper aims to highlight that biomedical image and signal processing are not exceptions to the pervasiveness of artificial intelligence across disciplines and fields. The subject's growing and widespread interest has sparked a massive study effort that is expanding at an exponential rate. Modeling, segmentation, registration, classification, and synthesis activities have been transformed by machine and deep learning models through analysis of vast and varied biological data, exceeding old methods. The authors believe that the data is not being fully utilized in the field, nevertheless, due to the challenges in turning them into information that may be used physiologically or therapeutically. By offering ways to make the models interpretable and by offering explanations, Explainable AI (XAI) aims to close this translational gap. The study suggests that - the community is becoming more interested in the many options that have been put forth thus far.

4. The study highlights deep learning's amazing achievement which has generated curiosity in its potential use in medical imaging diagnosis. Modern deep learning models can already classify several types of medical data with human-level accuracy, however due to their lack of interpretability, these models are rarely used in clinical workflows. The mysterious nature and the necessity to develop methods to explain these models' decision-making processes has increased as a result of deep learning models, giving rise to the concept of "eXplainable Artificial Intelligence" (XAI). In this context, the authors offer a comprehensive overview of XAI used for medical imaging diagnosis, encompassing visual, textual, example-based, and concept-based explanation techniques. In addition, their work reviews existing medical image datasets and existing metrics for evaluating image quality. Additionally, the authors added performance comparisons for different reporting methods. Furthermore, the main challenges in their application of XAI in medical image processing and future research directions on this topic are also discussed.

5. The study discusses the challenges of using ML when applying established methods for ensuring safety-critical systems and software because there isn't a clear, pre-defined specification against which to judge validity. The "opaque" character of this issue raises numerous difficulties. In order to address this problem, explainable AI (XAI) techniques have been proposed. These techniques create human-interpretable representations of ML models that can assist users in developing confidence and trust in the ML system. Yet, very little research specifically examines how explainability contributes to safety assurance in the context of ML development. This study outlines how XAI approaches can help ensure the safety of ML-based systems. The study then illustrates how XAI approaches can be used to create data to support safety assurance using a practical ML-based clinical decision support system about weaning patients from mechanical breathing. To demonstrate when and how XAI approaches can support a safety case, the results are also articulated in a safety argument. Overall, the authors draw the conclusion that while XAI approaches are important for ML-based systems in healthcare, they are insufficient on their own to guarantee safety.

6. The study presents that the impact of AI during the COVID-19 pandemic has been severely limited by the model's lack of transparency. This systematic review examines the use of explainable artificial intelligence (XAI) during the pandemic and how its use can overcome barriers to succeed in practice. The authors have found that using XAI effectively can improve model performance, increase end-user trust, and deliver the value needed to influence user decision-making. They have introduce the reader to common XAI techniques, their usefulness, and specific use cases. XAI performance assessment is also discussed as an important step to maximize the value of AI-based clinical decision support systems. The authors have illustrated classic, modern, and potential future XAI trends to explain the evolution of new XAI techniques. Finally, they present a checklist of suggestions during the experimental design process, supported by recent publications. Specific examples of potential solutions also address common challenges in implementing AI solutions. They hope this review can serve as a guide to improve the clinical impact of future AI-based solutions.

7. The study explains that "Background Eye" fungus and optical coherence tomography (OCT) scans can be used to quantify the rate at which different parts of the eye age; nevertheless, their genetic and environmental causes have proven difficult to identify. The authors put forward methods in order to capture two separate facets of eye (retinal, macula, and fovea) aging. They have trained convolutional neural networks to predict age from 175,000 eye fundus and OCT pictures from participants of the UK Biobank cohort. To detect unique genetic and environmental factors linked to the new age predictor, they conducted a genome-wide association study (GWAS) and high-throughput epidemiology, discovering factors linked to rapid eye aging.

8. The aim of the study is the development of a simple and interpretable Bayesian network (BN) for the classification of HPV status in patients with oropharyngeal carcinoma. This study included 246 patients, 216 of whom were HPV positive. The authors extracted 851 radiomic tracers from contrast-enhanced computed tomography (CT) images of the patients. The Mens eX Machina (MXM) method selected the two most significant predictors: sphericity and max2DDiameterRow. The area under the curve (AUC) showed the performance of the BN model in 30% of the test data. For comparison purposes, a method based on Support Vector Machine (SVM) was also implemented. The simple structure and interpretability of their BN model aimed to aid physicians in treatment decisions and enable non-invasive detection of HPV status from contrast-enhanced CT images.

9. The study aims to assess the current state of reproducibility and reusability of computational pathology algorithms. The authors have evaluated peer-reviewed articles available on Pubmed and published between January 2019 and March 2021, in five use cases: standardization of staining, type segmentation of tissue, feature evaluation, modification prediction, and direct extraction of classification, staging, and prediction information. In addition they developed criteria for the accessibility of data and code as well as the statistical analysis of the results and evaluated them in 161 publications. This review highlights candidates for reproducible and reusable algorithms in computational pathology. It is intended for both pathologists interested in deep learning and researchers applying deep learning algorithms to the challenges of computational pathology. The authors provide a list of reusable data tools and a detailed overview of the publications with their reproducibility and reusability criteria.

10. The study explains that DNA methylation has a significant impact on gene expression and can be associated with various diseases. Meta-analysis of available DNA methylation datasets requires the development of a dedicated pipeline for collaborative data processing. Therefore, the authors provide a comprehensive approach for combined DNA methylation datasets to classify controls and patients. The solution includes harmonizing data, building classification models for machine learning, reducing model dimensionality, imputing missing values, and explaining model predictions using Explainable Artificial Intelligence (XAI). They show that harmonization can improve classification accuracy by up to 20% when the preprocessing methods for training data and test datasets are different. Explainable AI approaches have allowed them to explain the model's predictions from both a population perspective and an individual perspective.

Bibliography

[1] Babak Haghighi, Warren B. Gefter, Lauren Pantalone, Despina Kontos, Eduardo Mortani Barbosa Jr, "Quantitative CT texture-based method to predict diagnosis and prognosis of fibrosing interstitial lung disease patterns", arXiv:2206.09766, Jun. 20, https://doi.org/10.48550/arXiv.2206.09766

[2] Zizhao Chen, Hongliang Wang, Chia-Hui Yeh, Xilin Liu, "Classify Respiratory Abnormality in Lung Sounds Using STFT and a Fine-Tuned ResNet18 Network", arXiv:2208.13943, Aug. 30, https://doi.org/10.48550/arXiv.2208.13943

[3] Guang Yang, Arvind Rao, Christine Fernandez-Maloigne, Vince Calhoun, Gloria Menegaz, "Explainable AI (XAI) in Biomedical Signal and Image Processing: Promises and Challenges", arXiv:2207.04295, Jul. 9, https://doi.org/10.48550/arXiv.2207.04295

[4] Cristiano Patrício, João C. Neves, Luís F. Teixeira, "Explainable Deep Learning Methods in Medical Imaging Diagnosis: A Survey", arXiv:2205.04766, May 10, https://doi.org/10.48550/arXiv.2205.04766

[5] Yan Jia, John McDermid, Tom Lawton, Ibrahim Habli, "The Role of Explainability in Assuring Safety of Machine Learning in Healthcare", arXiv:2109.00520, [v2] May 5, https://doi.org/10.48550/arXiv.2109.00520

[6] Felipe Giuste, Wenqi Shi, Yuanda Zhu, Tarun Naren, Monica Isgut, Ying Sha, Li Tong, Mitali Gupte, May D. Wang, "Explainable Artificial Intelligence Methods in Combating Pandemics: A Systematic Review", arXiv:2112.12705, [v4] Jun. 23, https://doi.org/10.48550/arXiv.2112.12705

[7] Alan Le Goallec, Samuel Diai, Sasha Collin, Théo Vincent, Chirag J. Patel, "Deep learning of fundus and optical coherence tomography images enables identification of diverse genetic and environmental factors associated with eye aging", Aug. 1, https://doi.org/10.1101/2021.06.24.21259471

[8] Oya Altinok, Albert Guvenis, "Interpretable Radiomics Method for Predicting Human Papillomavirus Status in Oropharyngeal Cancer using Bayesian Networks", Jun. 30, https://doi.org/10.1101/2022.06.29.22276890

[9] Sophia J. Wagner, Christian Matek, Sayedali Shetab Boushehri, Melanie Boxberg, Lorenz Lamm, Ario Sadafi, Dominik J. E. Waibel, Carsten Marr, Tingying Peng, "Built to last? Reproducibility and Reusability of Deep Learning Algorithms in Computational Pathology", May 31, https://doi.org/10.1101/2022.05.15.22275108

[10] Alena Kalyakulina, Igor Yusipov, Maria Giulia Bacalini, Claudio Franceschi, Maria Vedunova,  Mikhail Ivanchenko, "Disease classification for whole blood DNA methylation: meta-analysis, missing values imputation, and XAI", May 23, https://doi.org/10.1101/2022.05.10.491404