An XAI Approach to Deep Learning Models in the Detection of DCIS
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "An XAI Approach to Deep Learning Models in the Detection of DCIS".
Dev: Here's the gist of the paper. Deep Learning models have been employed over the past decade to improve the detection of conditions relative to the human body and in relation to breast cancer particularly <ref:2106.14186#pg2>.
Rosa: Our contribution attempts to interpret the early detection of breast cancer while enhancing clinicians’ confidence in such techniques through the use of eXplainable AI <ref:2106.14186#pg6>.
Dev: Research Focus and Methodology The study researched the best way to back-propagate a selected CNN model, previously developed in 2017; and adapted in 2019 <ref:2106.14186#pg3>.
Rosa: The methodology proved that it is possible to uncover the intricacies involved within a model; at neuron level, in converging towards the classification of a mammogram <ref:2106.14186#pg3>.
Title and authors: Rosa: Okay, so before we get into what they did, let's just nail down who wrote this. The paper is An XAI Approach to Deep Learning Models in the Detection of DCIS, and it’s written by Michele La Ferla, Matthew Montebello, and Dylan Seychell from the University of Malta.
Dev: Yeah, so it’s a team effort from researchers at that university focusing on this interpretability issue. It sets up a framework for how we can look inside these complex models to see the reasoning behind their output.
Taro: I wonder what their specific background is in the clinical side versus the deep learning side, because that usually dictates how useful the interpretation ends up being in a real setting.
Rosa: Well, they’re researchers tackling this from both sides—the clinical need for trust and the technical challenge of back-propagation—which should give them a solid grounding for what they are proposing.
Dev: It’s about bridging that gap between the complex math running on a computer and the practical decisions made by a radiologist looking at an image. That's the core goal of this work, connecting their technical approach to real medical practice.
The paper's summary: Rosa: So, what they actually did is test five different back-propagation methods on a CNN model that had been developed in two thousand seventeen and then updated in two thousand nineteen <ref:2106.14186#pg1>. They were trying to see which method was the best for uncovering the inner workings of the model at a neuron level when classifying a mammogram.
Dev: And they found that two techniques, Deep Taylor Decomposition and LRP-Epsilon, gave them the best results across their tests on a subset of twenty mammograms from their CBIS-DDSM dataset <ref:2106.14186#pg2>.
Taro: So they're comparing different mathematical tools for explaining what the model is doing when it looks at an actual mammogram image. That’s a very concrete comparison to make, moving away from just saying "the model is good."
Rosa: Right, and they noted that Deep Taylor Decomposition often performed better than the other three methods they tested, although LRP-Epsilon showed some better results in a couple of specific cases.
Dev: They also mentioned that this was done using the CBIS-DDSM dataset, which is an important detail because it’s a scanned film mammographic dataset that’s been annotated by expert radiologists and used widely in the deep learning community.
The paper's improvements: Rosa: Now, moving to what they suggest as improvements, they point out some real feedback from the specialists they interviewed. They noted that while this model approach is theoretically sound, the dataset itself could be better.
Dev: And that’s a big one because it means the quality of the input data matters a ton for how good any AI interpretation will actually be in practice, right?
Taro: I think they also flagged something specific about that dataset—they said it only displays a one-sided full-size mammogram for each case, which might limit what kind of information is available to the model.
Rosa: And then they suggested a combination: using a ResNet fifty architecture model alongside the Deep Taylor Decomposition method as a theoretically strong pairing for building deep learning models and applying backpropagation to them.
Dev: So the improvement isn't just one trick; it’s about combining a specific neural network structure with that specific mathematical decomposition technique to get this better interpretation. That combination seems like the most robust theoretical path they found.
Conclusion: Rosa: To wrap things up, these findings show that knowing how a model classified different mammograms—whether it's benign, malignant, or no signs of a tumour—is genuinely useful for radiologists who are trying to diagnose breast cancer.
Dev: They also pointed out that while this approach is promising, they need to improve the dataset because they felt the current one isn't as good as it could be for fully evaluating this XAI model.
Taro: And from my perspective, what their suggestion about choosing particular mammograms from the dataset and running the XAI model against them is a good practical next step to test its real-world utility before moving further.
Rosa: So, in short, they’ve shown that using techniques like Deep Taylor Decomposition on CNNs can reveal which neurons are important for classification, but they need better data and more specific testing scenarios to really make this tool reliable for clinical use.
Dev: It’s a solid piece of work because it takes the abstract idea of model interpretation and applies concrete mathematical methods to a high-stakes medical imaging task.
Michele La Ferla, Matthew Montebello, Dylan Seychell
University of Malta
eess.IV, cs.AI, cs.CV, cs.LG
Submitted: 2021-06-27
Updated: 2023-10-28
Comments: 12 pages, 5 figures
Journal ref: AIAI 2023. IFIP Adv.Inf.Comm.Tech. 677 (2023) 409-420
DOI: 10.1007/978-3-031-34171-7_33
Code: https://github.com/lishen/end2end-all-conv
Project page: http://www.eng.usf.edu/cvprg/Mammography/Database.html
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 62/100
The gist: Deep Learning models have been employed over the past decade to improve the detection of conditions relative to the human body and in relation to breast cancer particularly <ref:2106.14186#pg2>.
Key concepts
- Back-propagation
- This is a method used to train neural networks by adjusting the model's internal parameters based on errors. In this study, it was adapted to uncover how a selected CNN classifies a mammogram by tracing the influence back through its layers, revealing which neurons are most important for the final decision.
- Deep Taylor Decomposition
- This technique is used to approximate the relevance propagation function by breaking down complex functions into their partial derivatives. It helps researchers understand which specific neurons in a deep learning model contribute most significantly to classifying an image, particularly useful for images like mammograms.
- XAI Techniques (e.g., LRP-Epsilon)
- These are methods designed to explain the decisions made by complex AI models. They generate visualizations or maps that highlight which parts of the input data (like specific pixels in a mammogram) were most influential in the model's output, helping clinicians trust the diagnostic predictions.
- ResNet-50 Architecture
- This is a specific type of deep neural network architecture used in this research. ResNet-50 is known for its ability to learn complex features from images, and combining it with interpretation methods like Deep Taylor Decomposition provides a strong framework for analyzing how the model processes mammogram data.
Terminology
Summary
Deep Learning models have been employed over the past decade to improve the detection of conditions relative to the human body and in relation to breast cancer particularly <ref:2106. p. 0 mt
Abstract. Deep Learning models have been employed over the past
decade to improve the detection of conditions relative to the human body
and in relation to breast cancer particularly. However, their application
to the clinical domain has been limited even though they improved the
detection of breast cancer in women at an early stage. Our contribution
attempts to interpret the early detection of breast cancer while enhancing
clinicians’ confidence in such techniques through the use of eXplainable
AI.Where the paper says it
Where the paper says it
Research Focus and Methodology
The study researched the best way to back-propagate a selected CNN model, previously developed in 2017; and adapted in 2019 <ref:2106.14186#pg3>. The methodology proved that it is possible to uncover the intricacies involved within a model; at neuron level, in converging towards the classification of a mammogram <ref:2106. p. 0 We researched the best way to back-propagate a selected CNN model,
previously developed in 2017; and adapted in 2019. Our methodology
proved that it is possible to uncover the intricacies involved within a
model; at neuron level, in converging towards the classification of a mammogram. Our methodology
proved that it is possible to uncover the intricacies involved within a
model; at neuron level, in converging towards the classification of a mammogram. After conducting a number of tests using five back-propagation
methods, we noted that the Deep Taylor Decomposition and the LRPEpsilon techniques produced the best results. After conducting a number of tests using five back-propagation
methods, we noted that the Deep Taylor Decomposition and the LRPEpsilon techniques produced the best results. These were obtained on
a subset of 20 mammograms chosen at random from the CBIS-DDSM
dataset.Where the paper says it
Where the paper says it
Where the paper says it
Literature Context
Breast carcinoma is the most common types of tumours among women in the western world, according to WHO reports <ref:2106.14186#pg4>. The CBIS-DDSM scanned film mammographic dataset was selected due to its unique features, being that is has been carefully annotated by expert radiologists and has also been extensively used in the deep learning community <ref:2106. p. 1 This was followed by a comparative study between different available datasets
to adopt as part of our scientific tests that employed a Convolutional Neural
Network (CNN). The CBIS-DDSM scanned film mammographic dataset was
selected due to its unique features, being that is has been carefully annotated
by expert radiologists and has also been extensively used in the deep learning
community. The importance of LYNA was
however proved in two other studies carried out in 2018 and 2019, respectively
[5,16]. During the NIPS17 conference, one particular study made by Shen et al. During the NIPS17 conference, one particular study made by Shen et al. contributed significantly to tumour identification using screening mammogram
images [13].Where the paper says it
Where the paper says it
Where the paper says it
XAI Techniques Compared
The five chosen techniques used for comparison were: Deconvolution Network [1], Deep Taylor Decomposition [9], LRP-Preset A Flat [1], LRP-Epsilon [1], and Guided Backprop [1] <ref:2106.14186#pg4>. The results showed that the Deep Taylor Decomposition did in most cases perform better than the other four, however, there were instances where the LRP-Epsilon technique gave better results <ref:2106. p. 5 La Ferla
4 Results and Discussion
When comparing the five different methodologies we tested to determine the one
which interpreted mammograms best, we noticed that the Deep Taylor Decomposition did in most cases perform better than the other four, however, there
were instances where the LRP-Epsilon technique gave better results. In the following section, we will go through the different results achieved on the sample
subset of 20 mammograms from the CBIS-DDSM dataset. The figure shows the results obtained by the Deep Taylor
Decomposition method in the first column and the respective result obtained by
the Epsilon method to the right. For case 009, The Epsilon results show a larger
area which could contribute to the classification of the mammogram.Where the paper says it
Where the paper says it
Evaluation and Clinical Feedback
The initial comments received from the interviewed specialists was that the chosen dataset could have been a better one <ref:2106.14186#pg10>. The second point which the clinicians noted was the quality of the mammograms in the CBIS-DDSM dataset <ref:2106. p. 9 amount of detail which the BI-RADS mammograms go into can therefore be
much more than the ones processed by DICOM machines, therefore allowing for
microcalcifications in the breast to be identified by radiologists in an easier way. The second point which the clinicians noted was the quality of the mammograms in the CBIS-DDSM dataset. In particular case numbers: 072, 056, 063,
148 and 259 were ruled as being unidentifiable, and stating that if they were presented with such mammograms, they would ask for the radiographer to perform
a second mammographic test on the patient. A third point which the clinicians mentioned in the interview was that the
CBIS-DDSM dataset only displays a one-sided full-size mammogram of each
case. Our methodology
proved that it is possible to uncover the intricacies involved within a
model; at neuron level, in converging towards the classification of a mammogram. After conducting a number of tests using five back-propagation
methods, we noted that the Deep Taylor Decomposition and the LRPEpsilon techniques produced the best results.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Future Work
Knowing how a model has classified different mammograms into containing benign, malignant or no signs of a tumour would be beneficial to radiologists who are attempting to diagnose the presence of breast cancer <ref:2106.14186#pg8>. Improvements need to be made to the dataset for a better evaluation of the XAI model <ref:2106. p. 8 This
time, the response received involved less participants, but the study was more
focused and detailed such that we decided to meet three specialists in breast
radiology, and two in breast surgery individually to get a first-hand evaluation
from them on the proposed approach and model. The initial comments we received from the interviewed specialists was that
the chosen dataset could have been a better one. The CBIS-DDSM scanned film mammographic dataset was
selected due to its unique features, being that is has been carefully annotated
by expert radiologists and has also been extensively used in the deep learning
community. Additionally, we used this dataset on an already trained model, due
to a lack of computational resources, and to further focus our study on the
element of XAI in deep learning. There were
some instances in the subset, where the Guided Backprop method gave certain
information about regions in the mammogram where abnormality may occur;
such as in 016, 223 and 242; where there is a more dense amount of pixels
around particular regions of the mammograms. However, having said that, we
found that other methods performed better when attempting to identify tumours
in specific regions of a mammogram. While the model took this decision, we still
concluded that these mammograms should be interpreted also by a human radiologist. 5 Evaluation
To assist in the evaluation of the developed XAI model, we created a second
questionnaire and shared it with the same 12 clinical specialists from Malta who
were chosen for the initial questionnaire mentioned in the introduction.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
Alber, M., Lapuschkin, S., Seegerer, P., Hägele, M., Schütt, K.T., Montavon, G. p. 3 f(z)j =
e
zj
Pc
t=1 e
zt
forj = 1, . 1 Layer-Wise Relevance Propagation
Following the research by Montavon and Binder [10,3], it was found that the
Layer-wise Relevance Propagation method is one of the most effective algorithms
used to explain and interpret decisions in deep learning networks. f(z)j =
e
zj
Pc
t=1 e
zt
forj = 1, . 1 Layer-Wise Relevance Propagation
Following the research by Montavon and Binder [10,3], it was found that the
Layer-wise Relevance Propagation method is one of the most effective algorithms
used to explain and interpret decisions in deep learning networks. f(z)j =
e
zj
Pc
t=1 e
zt
forj = 1, . 1 Layer-Wise Relevance Propagation
Following the research by Montavon and Binder [10,3], it was found that the
Layer-wise Relevance Propagation method is one of the most effective algorithms
used to explain and interpret decisions in deep learning networks. The best performing solution would be a combination of
these two techniques to achieve ideal results. 2 Literature Review
In the early 1990s, one of the first datasets on breast cancer was published on the
Machine Learning Repository site.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Ayhan, M, S., Kümmerle, L.B., Kühlewein, L., Inhoffen, W., Aliyeva, G., Ziemssen, F., Berens, P.: Clinical validation of saliency maps for understanding deep neural networks in ophthalmology. p. 0 The results showed that XAI can indeed be used as a proof
of concept to begin discussions on the implementation of assistive AI
systems within the clinical community. Keywords: Explainable Artificial Intelligence, Layer-wise Relevance Propagation, Deep Taylor Decomposition, LRP-Epsilon, Breast Cancer, Deep Learning,
Convolutional Neural Network
1 Introduction
Breast carcinoma is the most common types of tumours among women in the
western world, according to WHO reports [19]. The mammographic presence of breast cancer can initially be noted through one
of the following four methods: by discovering minor distortions of the breast
tissue; showing the presence of masses in the breast; through the presence of
non-asymmetrical breasts; or through the presence of microcalcifications. arXiv:2106.Where the paper says it
Where the paper says it
Binder, A., Montavon, G., Lapuschkin, S., Müller, K.R., Samek, W.: Layer-wise relevance propagation for neural networks with local renormalization layers. p. 3 f(z)j =
e
zj
Pc
t=1 e
zt
forj = 1, . 1 Layer-Wise Relevance Propagation
Following the research by Montavon and Binder [10,3], it was found that the
Layer-wise Relevance Propagation method is one of the most effective algorithms
used to explain and interpret decisions in deep learning networks. The results showed that XAI can indeed be used as a proof
of concept to begin discussions on the implementation of assistive AI
systems within the clinical community. Keywords: Explainable Artificial Intelligence, Layer-wise Relevance Propagation, Deep Taylor Decomposition, LRP-Epsilon, Breast Cancer, Deep Learning,
Convolutional Neural Network
1 Introduction
Breast carcinoma is the most common types of tumours among women in the
western world, according to WHO reports [19]. An XAI Approach to Deep Learning Models in
the Detection of DCIS
Michele La Ferla1[0000
−0002
−2230
−6305]
Matthew Montebello1[0000
−0002
−6177
−7548]
Dylan Seychell1[0000
−0002
−2377
−9833]
University of Malta, Msida MSD 2080, Malta
https://www. mt/ict/ai michele. An XAI Approach to Deep Learning Models in
the Detection of DCIS
Michele La Ferla1[0000
−0002
−2230
−6305]
Matthew Montebello1[0000
−0002
−6177
−7548]
Dylan Seychell1[0000
−0002
−2377
−9833]
University of Malta, Msida MSD 2080, Malta
https://www. mt/ict/ai michele. An XAI Approach to Deep Learning Models in
the Detection of DCIS
Michele La Ferla1[0000
−0002
−2230
−6305]
Matthew Montebello1[0000
−0002
−6177
−7548]
Dylan Seychell1[0000
−0002
−2377
−9833]
University of Malta, Msida MSD 2080, Malta
https://www. mt/ict/ai michele. The mammographic presence of breast cancer can initially be noted through one
of the following four methods: by discovering minor distortions of the breast
tissue; showing the presence of masses in the breast; through the presence of
non-asymmetrical breasts; or through the presence of microcalcifications. arXiv:2106.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Montavon, G., Samek, W., Müller, K.R.: Methods for interpreting and understanding deep neural networks. p. 5 In all instances the A-Flat method enlarged those pixels which showed an abnormality, so it would be beneficial as a
first step to identify those mammograms which show a sign of abnormality, but
these are also visible to the naked eye, thus not contributing that much to the
analysis. After having analysed and eliminated two of the five options using a visual
comparison made using the naked eye, we remained with the LRP Epsilon, DeConvNet and the Deep Taylor Decomposition methods, to help us understand
how the neural network classified the mammograms. It was still too early to determine whether such an algorithm improved the efficiency of the procedure or diagnostic accuracy. The importance of LYNA was
however proved in two other studies carried out in 2018 and 2019, respectively
[5,16]. The mammographic presence of breast cancer can initially be noted through one
of the following four methods: by discovering minor distortions of the breast
tissue; showing the presence of masses in the breast; through the presence of
non-asymmetrical breasts; or through the presence of microcalcifications. arXiv:2106.Where the paper says it
Where the paper says it
Where the paper says it
Reyes, M., Meier, R., Pereira, S., Silva, C., Dahlweid, M., Tengg-Kobligk, H., Summers, R., Wiest, R.: On the interpretability of artificial intelligence in radiology: Challenges and opportunities. p. 0 The results showed that XAI can indeed be used as a proof
of concept to begin discussions on the implementation of assistive AI
systems within the clinical community. Keywords: Explainable Artificial Intelligence, Layer-wise Relevance Propagation, Deep Taylor Decomposition, LRP-Epsilon, Breast Cancer, Deep Learning,
Convolutional Neural Network
1 Introduction
Breast carcinoma is the most common types of tumours among women in the
western world, according to WHO reports [19]. The results showed that XAI can indeed be used as a proof
of concept to begin discussions on the implementation of assistive AI
systems within the clinical community. Keywords: Explainable Artificial Intelligence, Layer-wise Relevance Propagation, Deep Taylor Decomposition, LRP-Epsilon, Breast Cancer, Deep Learning,
Convolutional Neural Network
1 Introduction
Breast carcinoma is the most common types of tumours among women in the
western world, according to WHO reports [19].Where the paper says it
Where the paper says it
Rose, C., Turi, D., Williams, A., Wolstencroft, K., Taylor, C. p. 0 Our methodology
proved that it is possible to uncover the intricacies involved within a
model; at neuron level, in converging towards the classification of a mammogram. After conducting a number of tests using five back-propagation
methods, we noted that the Deep Taylor Decomposition and the LRPEpsilon techniques produced the best results. An XAI Approach to Deep Learning Models in
the Detection of DCIS
Michele La Ferla1[0000
−0002
−2230
−6305]
Matthew Montebello1[0000
−0002
−6177
−7548]
Dylan Seychell1[0000
−0002
−2377
−9833]
University of Malta, Msida MSD 2080, Malta
https://www. mt/ict/ai michele. Shen et al. used the publicly available well-curated CBIS-DDSM and inBreast mammographic datasets of scanned film mammography studies to build
their model upon [12].Where the paper says it
Where the paper says it
Where the paper says it
Shen, L.: lishen/end2end-all-conv (2019), https://github. p. 0 An XAI Approach to Deep Learning Models in
the Detection of DCIS
Michele La Ferla1[0000
−0002
−2230
−6305]
Matthew Montebello1[0000
−0002
−6177
−7548]
Dylan Seychell1[0000
−0002
−2377
−9833]
University of Malta, Msida MSD 2080, Malta
https://www. mt/ict/ai michele.Where the paper says it
Street, W.N.: Cancer diagnosis and prognosis via linear-programming-based machine learning. p. 1 The best performing solution would be a combination of
these two techniques to achieve ideal results. 2 Literature Review
In the early 1990s, one of the first datasets on breast cancer was published on the
Machine Learning Repository site.Where the paper says it
Stumpe, M.: Assisting pathologists in detecting cancer with deep learning (2021), https://ai. p. 2 An XAI Approach to Deep Learning Models in the Detection of DCIS 3
which have portrayed biopsies from MRI stage procedures in the second stage of
a diagnosis and used to develop a deep learning model able to detect a tumour
which is at a metastasis stage [18]. While the LYNA model proved to be significantly more effective in detecting breast cancer, the researchers themselves
admitted in their study that an accurate model alone does not prove enough to
improve the diagnostic work done by pathologists or improve outcomes for breast
cancer patients [8]. An XAI Approach to Deep Learning Models in the Detection of DCIS 3
which have portrayed biopsies from MRI stage procedures in the second stage of
a diagnosis and used to develop a deep learning model able to detect a tumour
which is at a metastasis stage [18]. While the LYNA model proved to be significantly more effective in detecting breast cancer, the researchers themselves
admitted in their study that an accurate model alone does not prove enough to
improve the diagnostic work done by pathologists or improve outcomes for breast
cancer patients [8].Where the paper says it
Where the paper says it
WHO: Incidence of female breast cancer per 100000. Accessed September. 07, 2020. [Online] (2019), https://gateway. p. 0 An XAI Approach to Deep Learning Models in
the Detection of DCIS
Michele La Ferla1[0000
−0002
−2230
−6305]
Matthew Montebello1[0000
−0002
−6177
−7548]
Dylan Seychell1[0000
−0002
−2377
−9833]
University of Malta, Msida MSD 2080, Malta
https://www. mt/ict/ai michele.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Back-propagation model integration for CNN interpretation allows clinicians to
uncover the intricacies involved within a model; at neuron level, in converging towards the classification of a mammogram.
This enables clinicians to understand how a model classifies between benign, malignant, or non-tumorous breast tissue. -
Deep Taylor Decomposition can be used to
approximate the relevance propagation function
by decomposing the function into its partial derivatives. This method helps us understandwhich neurons which contribute to the classification of an image and work best with monochromatic images such as the ones used in mammograms.
-
The combination of a ResNet-50 architecture and Deep Taylor Decomposition is theoretically proven to be a good combination for building deep learning models and applying backpropagation. This hybrid system can provide a more accurate interpretation by leveraging the strengths of both methods.
Abstract
The results showed that XAI could indeed be used as a proof of concept to begin discussions on the implementation of assistive AI systems within the clinical community.
Related papers
- HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception
- A Disease-Centric Vision-Language Foundation Model for Precision Oncology in Kidney Cancer
- CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography
- Task-Aware Bitrate Thresholds for Codec-Robust Ultrasound Segmentation