<abstract xmlns="http://www.w3.org/1999/xhtml">

<p>Automatic detection of tissue types on whole-slide images (WSI) is an important task in computational histopathology that can be solved with convolutional neural networks (CNN) with high accuracy. However, the black-box nature of CNNs rightfully raises concerns about using them for this task. In this paper, we reformulate the task of tissue type detection to multiple binary classification problems to simplify the justification of model decisions. We propose an adapted Bag-of-local-Features interpretable CNN for solving this problem, which we train on eight newly introduced binary tissue classification datasets. The performance of the model is evaluated simultaneously with its decision-making process using logit heatmaps. Our model achieves better performance than its non-interpretable counterparts, while also being able to provide human-readable justification for decisions. Furthermore, the problem of data scarcity in computational histopathology is accounted for by using data augmentation techniques to improve both the performance and even the validity of model decisions. The source code and binary datasets can be accessed at: <ext-link ext-link-type="uri" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="https://github.com/galigergergo/BolFTissueDetect">https://github.com/galigergergo/BolFTissueDetect</ext-link>.</p>
</abstract>

Automatic detection of tissue types on whole-slide images (WSI) is an important task in computational histopathology that can be solved with convolutional neural networks (CNN) with high accuracy. However, the black-box nature of CNNs rightfully raises concerns about using them for this task. In this paper, we reformulate the task of tissue type detection to multiple binary classification problems to simplify the justification of model decisions. We propose an adapted Bag-of-local-Features interpretable CNN for solving this problem, which we train on eight newly introduced binary tissue classification datasets. The performance of the model is evaluated simultaneously with its decision-making process using logit heatmaps. Our model achieves better performance than its non-interpretable counterparts, while also being able to provide human-readable justification for decisions. Furthermore, the problem of data scarcity in computational histopathology is accounted for by using data augmentation techniques to improve both the performance and even the validity of model decisions. The source code and binary datasets can be accessed at: https://github.com/galigergergo/BolFTissueDetect.

Explainable patch-level histopathology tissue type detection with bag-of-local-features models and data augmentation

Acta Universitatis Sapientiae, Informatica

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

Automatic detection of tissue types on whole-slide images (WSI) is an important task in computational histopathology that can be solved with convolutional...

Explainable patch-level histopathology tissue type detection with bag-of-local-features models and data augmentation

Online veröffentlicht: 08. Aug. 2023

Seitenbereich: 60 - 80

Eingereicht: 09. Mai 2023

DOI: https://doi.org/10.2478/ausi-2023-0006

Schlüsselwörter
computational pathology, tissue type detection, explainable artificial intelligence, convolutional neural networks, data augmentation

© 2023 Gergő Galiger et al., published by Sciendo

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

Explainable patch-level histopathology tissue type detection with bag-of-local-features models and data augmentation

Online veröffentlicht: 08. Aug. 2023

Seitenbereich: 60 - 80

Eingereicht: 09. Mai 2023

DOI: https://doi.org/10.2478/ausi-2023-0006

Schlüsselwörtercomputational pathology, tissue type detection, explainable artificial intelligence, convolutional neural networks, data augmentation

© 2023 Gergő Galiger et al., published by Sciendo

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

Schlüsselwörter
computational pathology, tissue type detection, explainable artificial intelligence, convolutional neural networks, data augmentation