In the healthcare domain, the analysis of images and omics data by means of methods based on Artificial Intelligence (AI) is increasingly providing meaningful insights into a number of diseases. However, accurate diagnoses and efective treatments require, in general, complex and multimodal analyses of medical data; making fruitful use of such data, due to their complex, sensitive and heterogeneous nature, is not straightforward. In recent years, Machine Learning and Deep Learning techniques emerged as powerful tools to perform specific disease detection and classification. In this work, we describe the ongoing projects of our research group related to this field.
ity of data, Artificial Intelligence (AI) has become an With the advancement of technology and the availabil- of several diseases. In this context, novel applications essential tool for medical professionals to improve pa- ously from new streams of data without forgetting the to acquire the huge amounts of data needed for efectively tackling more general tasks, such as the diagnosis have been proposed to allow models to learn continupreviously learned knowledge. This is the principle of Continual Learning (CL), a novel paradigm by which the trained model is computed via data streams where tasks and data are only available over time. With this approach, the model can incrementally learn and autonomously [1]. Indeed, in medical contexts, the training from scratch of models that would be able to accomplish and predict a wide number of possible tasks and activities requires model can be trained again to solve new problems [2].
Considering the ability of DL to learn complex patterns, such methods have been particularly successful also in analyzing omics data. Omics analysis involves the study of biological molecules, including DNA, RNA, proteins, and metabolites. They can provide insights into the underlying molecular mechanisms of diseases and help identify biomarkers for diagnosis, prognosis, and personalized treatment. AI-based techniques, particularly DL, have been extensively applied in omics analysis, including genomics, transcriptomics, proteomics, and metabolomics. DL methods have been shown to provide accurate predictions of disease outcomes, discover novel biomarkers, and support clinicians in the proper treatment design. In the field of functional genomics, starting from the results of the Human Genome Project, the evo© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License lution of high-throughput and next-generation sequencing Indeed, by using such sequencing techniques it is pos- NNs predictions, leading to improved insights and better sible to measure the expression of thousands of genes decision-making. Another state-of-the-art XAI approach for each patient and hence to collect quantitative Gene for imaging is Gradient-weighted Class Activation MapExpression Profiles (GEP) to be used for research and ping (Grad-CAM) [8] which is a popular technique for clinical purposes. In recent years, DL has been widely visualizing and understanding decisions made by a CNN adopted in this field, providing breakthrough results and model. Specifically, it allows for visualizing the regions meaningful insights into the relationship between ge- of an input image that are most important for predicting nomics and cancer. A number of recent studies propose a particular class. In particular, Grad-CAM produces a and evaluate new approaches for feature selection (FS) heatmap that highlights the relevant regions in the input on GEP for cancer diagnosis and prognosis[3]. They aim image that the model used during the decision-making to select the most informative genes able to characterize process. classes and identify groups of patients. Despite the increasing number of DL-based applica
Although very promising, DL models are in general tions in medical imaging and omics analysis, there are not immediately interpretable, making it dificult to un- still challenges that need to be addressed, such as the derstand the causal relationship between the inputs and limited availability of high-quality annotated datasets, their outcomes. In the bioinformatics domain, this is an the interpretability of DL models, and the generalization even more severe problem. Models interpretability is of the models to new datasets. Therefore, our research indeed crucial to understand, for example, in the case activities aim at facing such challenges by designing innoof genomics, how the expression of a gene afects the vative solutions based on DL. In the following sections, progression of oncological disorders. we describe our research activities applied to medical
In this context, the adoption of Explainable Artificial imaging, in Section 2, and genomics-scale data analysis, Intelligence (XAI) methods has started to gain momen- in Section 3. Eventually, we discuss the future direction tum for interpretability purposes as well as to enhance and challenges for AI in the medical field and draw the FS [4, 5, 6]. conclusion in Section 4.
More in general, the black-box nature of Neural Networks (NNs) often limits the interpretability of their results. Advances in XAI provide various methods for inter- 2. AI for Medical Imaging preting black box models, ofering a clearer understanding of their predictions. For example, shapely Additive 2.1. Continual Learning for medical ex-Planation (SHAP) [7] is a game theory-based approach image classification for interpreting black box models. It assigns an impor- We proposed the use of CL approach to support DL architance value, called a SHAP value, to each feature, based tectures in classifying medical images [2]. In particular, on its contribution to the predictions. The SHAP method we used a collection of standardized biomedical photos, provides a way to understand the underlying workings of
MedMNIST v2 [9] that is composed of 12 datasets for 2D portant features in the classification of a particular disimages and 6 datasets for 3D. An example of the dataset ease, as shown in Figure 3. is shown in Figure 1. Once these regions are identified, they can be analyzed
We focused our experiments on the following data sets: to discover patterns in Chest X-ray images related to the Colon Pathology, Dermatoscope, Retinal OCT, Blood Cell disease in question.
Microscope, and Kidney Cortex Microscope. We used In addition to identifying patterns, we used these highand compared diferent CL strategies (i.e., Naive, Replay, lighted regions to analyze the correlation between these CWR*, ICaRL, Cumulative) [10]. areas and classification accuracy. This involves examin
Our results show that Replay strategies do not reach a ing the performance of the model when these regions good performance; this is probably due to the unbalanced are removed and comparing it to the performance when distribution of images across classes and a high number the full image is used. If the removal of these regions reof images. sults in a significant reduction in classification accuracy,
Instead, we found promising results for both CWR* and it suggests that these regions are indeed important for ICaRL according to the accuracy and forgetting value. accurate disease diagnosis.
In particular, in our experiments, we performed both 2.2. XAI in Chest X-ray classification quantitative and qualitative analyses to confirm that the portions identified by Grad-CAM are actually significant.
We investigated the use of CNNs to perform multiple- As for a quantitative assessment, a substantial decrease of disease classification from Chest X-ray images [ 11]. Dis- Recall after the removal (on average around 5%) is shown. eases that are a matter of concern for our experiments are As for a qualitative assessment, we took advantage from COVID-19 and tuberculosis (TB) Pneumonia. In partic- the TB dataset, which features labels provided by exular, we achieved promising results in the classification pert clinicians, representing the approximate location of task (i.e., Recall mean value: 0.89). the abnormality in the lung. In many cases, Grad-CAM
Furthermore, we analyzed the CNNs-based model to highlights the same areas suggested by clinicians. identify the mechanisms and the motivations steering NN decisions in the classification task. Specifically, we 2.3. Segmentation used Grad-CAM to identify the image areas used by the model to take a particular decision. The Grad-CAM tech- Semantic segmentation deals with labeling each pixel nique uses the gradient information flowing into the final on the image with a specific class. In the medical field, convolutional layer of a NN to compute the importance semantic segmentation represents a huge breakthrough of each feature map. It then weights the activation maps and it has been widely used for the detection of tumors, of the final convolutional layer based on the importance identifying diferent organs, and classifying specific tisof each feature map and averages them to produce a sue [12]. heatmap that highlights the important regions in the input image.
Then, the highlighted regions represent the most im2.3.1. Laryngeal Endoscopic Images number of patients (samples); the second problem concerns imbalanced classes: in most of the cases, there is a We proposed a DL-based approach to perform semantic significant diference between the number of instances segmentation of Laryngeal Endoscopic Images [13, 14]. in each group of interest. Finally, sequencing data are The dataset consists of 536 manually segmented in vivo typically collected from multiple sources, diferent labcolor images (512×512 pixels) of the larynx captured dur- oratories, and sequencing tools. This results in noisy ing two diferent resection surgeries. In particular, im- datasets which are dificult to analyze [ 15]. ages are categorized in 7 classes: void, vocal folds, other We proposed a new algorithm for genomic-scale analysis, tissue, glottal space, pathology, surgical tool and intubation. based on DL and XAI, whose aim is threefold: first, select In order to improve the quality of our prediction, we took the most meaningful genes for a regression/classificaadvantage of the potential coming from the declarative tion problem; second, provide a more accurate prediction nature of rule-based languages such as Answer Set Pro- model; third, quantify and evaluate the feature’s contribugramming (ASP). In particular, we included ASP in the tion to the predictions through XAI [16]. The proposed training to drive NN decisions and penalize misclassifi- algorithm is based on two main ideas: (1) recognize simication according to prior knowledge. Also, we refined larly correlated features using clustered correlation mathe prediction via rule-based methods, removing noise trix and then filter the redundant information for each (i.e., small “islands” of misclassified pixels) and wrong group by using Autoencoders (AEs). In contrast with predicted classes (i.e., classes that do not respect medical previous works, where AEs are used for dimensionalrequirements). ity reduction [17], we implemented a mechanism to still
Figures 2 show a visual example of the results achieved work at the level of the original features. We hence proby U-Net on two images. vide a more treatable dataset in terms of dimensionality,
In general, our approach achieved a mean IoU value without afecting interpretability; (2) we train NNs and higher than 0.7, with a relevant improvement after post- we iteratively select the most meaningful features using a processing. new ad-hoc defined XAI score. We eventually use the set of selected features (from all the iterations) to train and 3. AI for Omics Data Analysis explain a final model. The proposed algorithm workflow is shown in 4.
Functional genomics data, and in particular GEP datasets, We used our algorithm for the GEP analysis of Chronic represent a valuable source of information in medicine: Lymphocytic Leukemia (CLL) patients. The dataset was they are indeed used for diagnosis, prevention, and pre- composed of the GEP of 97 patients for 19367 genes. For cision medicine. However, their analysis results are chal- each patient a real-valued number was provided, indilenging for three main reasons. The first one is the course cating, as a factor of prognosis, the time interval after of dimensionality: a genomics dataset typically consists which the condition of the patient deteriorates. We disof a very large number of genes (features) for a small tinguished two classes: a first one for patients whose A Final model is trained and explained using al the N genes selected during iterations.
prediction Number of genes ≥ N Col ect selected genes Number of genes < N
Explainability-based Selection Select the most meaningful genes according to an ad-hoc defined
SHAP-based score
Correlation Clustering
Autoencoder (AE) Filter Clusters k1 . . . kq
Patients used a1sfea.tu.re.sforneach k1 . . . kq
AE.1.
. AEq Remove selected genes from k1 . . . kq condition deteriorates in a period shorter than 24 months, nary study will evolve into a large-scale testing campaign and a second class for the patients whose condition de- to assess the algorithm’s performance on a wide set of teriorates in a period equal to or longer than 24 months. state-of-the-art genomics datasets. We used the proposed algorithm for training a NN to solve such a classification problem as well as to identify a set of meaningful genes over the whole set of 19367. 4. Conclusion Such selected genes, together with their ad-hoc defined AI has been playing an increasingly important role in XAI-score are reported in Figure 5. the medicine and healthcare domain over the past few In a more general context, the proposed algorithm can be years, particularly in the areas of medical imaging and suitably used as a tool in genomics to identify protective omics data analysis. Nevertheless, new techniques and (or not) sets of genes for a disease, suggesting potential methods are needed for properly facing complex issues, pathways for further medical investigation. Our prelimi- such as scarce or low-quality data, yet improving diagnos
This work has been partially funded by PON “Ricerca e Innovazione” 2014-2020, CUP: H25F21001230004. tic accuracy, disease prediction, and treatment planning, leading to better patient outcomes and improved quality of life. However, the integration of AI into clinical practice must be carefully managed to ensure that these technologies are used ethically, responsibly, and safely.
Ongoing research and development in this field will be critical to realizing the full potential of AI in medicine and improving patient care. from deep networks via gradient-based localization, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 618–626. [9] J. Yang, R. Shi, D. Wei, Z. Liu, L. Zhao, B. Ke, H. Pfister, B. Ni, Medmnist v2: A large-scale lightweight benchmark for 2d and 3d biomedical image classification, arXiv preprint arXiv:2110.14795 (2021). [10] V. Lomonaco, L. Pellegrini, A. Cossu, A. Carta,
G. Grafieti, T. L. Hayes, M. De Lange, M. Masana, J. Pomponi, G. M. Van de Ven, et al., Avalanche: an end-to-end library for continual learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp.
3600–3610. [11] P. Bruno, F. Calimeri, Understanding automatic pneumonia classification using chest x-ray images, in: AIxIA 2020–Advances in Artificial Intelligence: [1] C. S. Lee, A. Y. Lee, Clinical applications of contin- XIXth International Conference of the Italian Asual learning machine learning, The Lancet Digital sociation for Artificial Intelligence, Virtual Event, Health 2 (2020) e279–e281. November 25–27, 2020, Revised Selected Papers, [2] A. Quarta, P. Bruno, F. Calimeri, Continual learn- Springer, 2021, pp. 37–50.
ing for medical image classification, in: F. Calimeri, [12] T. Dhamija, A. Gupta, S. Gupta, R. Katarya, G. Singh, M. Dragoni, F. Stella (Eds.), 1st AIxIA Workshop on Semantic segmentation in medical images through Artificial Intelligence For Healthcare (HC@AIxIA transfused convolution and transformer networks, 2022) co-located with the 21st International Con- Applied Intelligence 53 (2023) 1132–1148. ference of the Italian Association for Artificial In- [13] M.-H. Laves, J. Bicker, L. A. Kahrs, T. Ortmaier, telligence (AIxIA 2022), Udine, Italy, November A dataset of laryngeal endoscopic images with 30 2022, volume 3307 of CEUR Workshop Proceed- comparative study on convolution neural networkings, CEUR-WS.org, 2022, pp. 67–76. URL: https: based semantic segmentation, International jour//ceur-ws.org/Vol-3307/paper7.pdf. nal of computer assisted radiology and surgery 14 [3] E. Alhenawi, R. Al-Sayyed, A. Hudaib, S. Mirjalili, (2019) 483–492.
Feature selection methods on gene expression mi- [14] P. Bruno, F. Calimeri, C. Marte, M. Manna, Comcroarray data for cancer classification: A systematic bining deep learning and asp-based models for the review, Computers in Biology and Medicine 140 semantic segmentation of medical images, in: Rules (2022) 105051. and Reasoning: 5th International Joint Conference, [4] G. Graham, N. Csicsery, E. Stasiowski, G. Thou- RuleML+ RR 2021, Leuven, Belgium, September venin, W. H. Mather, M. Ferry, S. Cookson, J. Hasty, 13–15, 2021, Proceedings 5, Springer, 2021, pp. Genome-scale transcriptional dynamics and envi- 95–110. ronmental biosensing, Proceedings of the National [15] L. Koumakis, Deep learning models in genomics; Academy of Sciences 117 (2020). are we there yet?, Computational and Structural [5] J. Meena, Y. Hasija, Application of explainable arti- Biotechnology Journal 18 (2020) 1466–1473. ifcial intelligence in the identification of squamous [16] C. Adornetto, G. Greco, A new deep learning and cell carcinoma biomarkers, Computers in Biology xai-based algorithm for features selection in geand Medicine 146 (2022) 105505. nomics, 2023. a r X i v : 2 3 0 3 . 1 6 9 1 4 . [6] M. R. Karim, M. Cochez, O. Beyan, S. Decker, [17] P. Danaee, R. Ghaeini, D. A. Hendrix, A deep learnC. Lange, Onconetexplainer: explainable predic- ing approach for cancer detection and relevant gene tions of cancer types based on gene expression data, identification, in: Pacific symposium on biocomin: 2019 IEEE 19th International Conference on puting 2017, World Scientific, 2017, pp. 219–229. Bioinformatics and Bioengineering (BIBE), IEEE, 2019, pp. 415–422. [7] S. M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions, Advances in neural information processing systems 30 (2017). [8] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam,
D. Parikh, D. Batra, Grad-cam: Visual explanations