The use of deep learning in biomedical imaging and omics data has shown great potential for enhancing medical diagnosis and improving patient outcomes. In this paper, we present the deep learning and machine learning research activities of two research laboratories: EIDOS and qBio. Our research encompasses a broad range of topics, including digital pathology, integration of omics data, digital radiology, computational epidemiology and neuroimaging. We collaborate with several hospitals for the collection of relevant datasets and with international research centers and foreign universities to develop state-of-the-art techniques. Overall, we believe that the activities of these laboratories in deep and machine learning have the potential to improve the way we diagnose and treat various medical conditions.
committee of the Bioinformatics Italian Society (BITS)
[
The experience of EIDOS in biomedical image
processing and analysis is grounded in several projects and
collaboration with major hospitals. EIDOS was recently
supported by the EU through the DeepHealth project [
for Health and currently by Regione Piemonte through
volved. For instance, in the ONCOBIOME project [
cancer incidence, prognosis and prediction of treatment response funded by the EC under the topic SC1-BHC03-2018, the qBio group provided the analysis and integration of omics data (metagenome, small noncoding RNAomics, transcriptomic, metabolomics) by machine learning techniques to identify diagnostic and prognostic biomarker in four solid cancers. In the Minimal residual disease in follicular and mantle cell lymphoma: development and validation of novel tools and predictive models project funded by the Italian Ministry of Health (Finalizzata 2021) the qBio group developed an innovative approach for clustering functional data to post-treatment outcome predictor in blood cancers. Diferently in the
on the NextGeneration EU-funded Recovery and Resilience National Plan (PNRR) – Research Infrastructure to support the Italian network of collections of microorganisms the qBio group works on the implementation of ties of the project. Finally, in TrustAlert project, founded by ”La Compagnia di San Paolo” and ”Fondazione CDP” (grant dedicated to AI) the qBio group works on developing a smart platform for providing early warnings, monitoring, and forecasting tools for public health response agencies and local healthcare services for anticipating medical needs. 2. Digital pathology standard DL classification pipelines not straightforward due to computational requirements. Also, when dealing with H&E-stained images, it is often necessary to account for the diferent concentrations of the two histological stains, depending on how the staining procedure was carried on. This is why we also developed torchstain a popular stain normalization tool aimed at DL applications [14] which has increasingly gained traction in the community. 2.2. Generative models for synthetic data augmentation Our group has been actively engaged in exploring the potential of deep learning techniques for digital pathology applications. In this section, we present an overview of our group activities in this domain, focusing on three key areas of research. Firstly, we discuss our eforts in creating a large dataset of whole-slide images of colorectal polyps for diagnostic purposes. Secondly, we delve into our work on generative models for synthetic data augmentation to enhance the accuracy and robustness of our models. Finally, we describe our eforts in developing deep learning-based methods for grading colorectal cancer. Our goal is to provide a comprehensive overview of our group’s contributions in this field and to highlight the potential of deep learning for advancing digital pathology.
due to the underrepresented high-grade class in the UniToPatho dataset. To address this, we explored the use of Generative Adversarial Networks (GANs) to generate new samples for augmentation. In our initial experiment, we trained StyleGANs on diferent resolutions using both unconditional (one GAN for each class) and conditional (by conditioning on the grade) settings. Our analysis revealed intriguing details of the Generator network, such as high-quality generations almost indistinguishable from real ones, as well as latent-space properties that determine the distribution and positioning of cells in the output images. Surprisingly, we discovered that, with the same method described in [12], training a ResNet for 2.1. The UnitoPatho data collection classification solely on synthetic data produced nearly Digital histopathology solutions have gained increasing the same results as the real-data approach, with just 1.5% demand, fuelued by the widespread adoption of cancer lower accuracy on the real test set. However, despite screening programs [9]. In particular we have been chal- augmenting the real dataset with multiple variants of lenged by gastrointestinal histopathologists, who inspect synthetic samples, we did not significantly improve the tissue samples collected during colonoscopies, to pro- classification accuracy of the ResNets. The strong imbalvide automatic tools to recognize and classify colorectal ance in the dataset hindered our GANs from learning the polyps. Colorectal polyps are pre-malignant lesions that real distribution of the high-grade class. To overcome are analyzed to i) classify the polyp type (hyperplas- this, we developed a novel StyleGAN architecture that tic, adenoma) and ii) to evaluate the dysplasia grade in guides the synthesis process using high-level tissue feacase of adenomas. In this field the search for computer tures defined through segmentation masks of nuclei. We aided solutions is of paramount importance not only for created a new GAN based on the initial StyleGAN that the common need to simplify and speed-up the pathol- uses an additional UNet-like architecture for injecting ogists’ clinical routine. Indeed, the concordance rate the segmentation masks into the synthesis network. Our among pathologists is dificult to guarantee: for instance, new model achieved competitive FID, comparable with the concordance in assessing a tubulo-villous polyp or our initial GAN approach. Additionally, we built a tool low grade dysplasia is reported to be around 70% [10]. that allows editing segmentation masks and visualizing From the technical point of view, the nature of this task the generated results in real time. We plan to use these poses a number of challenges that must be taken into novel models and tools for a more precise augmentaaccount, which have been described in our recent pub- tion of the underrepresented class, by exploiting expert lications [11, 12]. The first issue to overcome, as usual, medical knowledge when manipulating the nuclei masks. is the scarce availability of data. We tackled this issue by building and releasing UniToPatho [12, 13], a high- 2.3. Colorectal cancer grading resolution annotated dataset of Hematoxylin and Eosin integrating heterogeneous features (H&E)-stained colorectal patches extracted from wholeslide images (WSI). Another challenge is posed by the huge resolution of WSIs, which makes the application of At the basis of the personalized medicine approaches for the prediction of stage and for the prevention of the disease there are new AI computational approaches based on the exploitation and integration of data from omics stud- scribed in [16, 17]. Our collaboration with the radiology ies. The identification of patients with a poor prognosis, units of Città della Salute e Della Scienza (CDSS), Maunonresponders to standard therapies, or with an elevated riziano, San Luigi, Monzino and ASLTo3 have evolved probability to have an adverse efect is the main goal of into the regionally funded C.o.R.S.A. project1. The joint the integrative omics data approach. The identification efort on this topic led to the CORDA data collection, of a heterogeneous signature based on transcriptomics, which is publicly available for download2, and contains metabolomics, metagenomics data, and the imaging fea- around 3000 images from patient who underwent Covidtures extracted from cancer slides could be used as pre- 19 screening, along with the ground-truth label obtained dictive of the progression of cancer and therefore adapt with RT-PCR testing (swab). Our ongoing eforts focus preventive therapies. The signature identification is gen- on obtaining models which are robust to biases in the erally obtained by the exploration of both early and late data [18], robust to noise given by diferent acquisition integration strategy [15]. The derived signatures will be sites and which can provide some form of explainabilbuilt either on a combination of individual features (such ity. For the latter, in [17] a DL diagnostic approach that as expression level or abundance/presence of a metabo- imitates the radiologist diagnosis process is proposed, lite) or of composite features - summary values of groups based on a preliminary classification stage mapping onto of highly correlated variables. standard radiological findings from the lungs, on top of which the Covid-19 is diagnosed.
In recent years, digital radiology has seen a surge in research and development due to the advancements in machine learning and computer vision techniques. Our group has been actively involved in exploring various aspects of digital radiology, ranging from Covid-19 detection to lung nodule detection and calcium score prediction. In this section, we will discuss our group’s activities and contributions in these areas, highlighting the C.o.R.S.A. Project for Covid-19 detection, the UnitoChest dataset for lung nodule detection, and our eforts in calcium score prediction from CXR. These projects represents a significant step forward in the field of digital radiology and have the potential to improve patient outcomes by enabling earlier and more accurate diagnoses. 3.1. COVID-19 and the C.o.R.S.A. project 3.2. Lung nodules segmentation and the
UnitoChest dataset Lung cancer is the primary cause of death for men and women, with a survival rate lower than breast and prostate cancer [19]. Therefore, early detection of lung nodules is the key to early cancer diagnosis and treatment efectiveness assessment. Deep neural networks achieve outstanding lung nodules detection, classification, and segmentation results. However, the quality and quantity of the training images can boost their performance. Within the DeepHealth project, we created UniToChest [20], a dataset consisting Computed Tomography (CT) scans of 623 patients. UniToChest is publicly available3 and is the largest of its kind and boasts a diversity of patient ages, acquisition machines and nodules diameters. Manual lung nodules segmentation is timeconsuming and prone to errors; so, several systems based on deep learning have been proposed for the detection and segmentation of lung nodules. In our study [21], we analyzed a U-Net based architecture that yields promising results in both detection and segmentation of lung nodules. Future research directions of this work include exploiting the three-dimensional information of nodules across neighboring slices.
At the peak of the Covid-19 pandemic in Italy in 2020, our group started a collaboration with the radiology units of local hospitals to help with the screening of Covid-19 patients. The first results of our efort toward Covid19 detection from chest X-ray (CXR) can be found in [16, 17]. Beside the crude achievements in terms of detection performance (sensitivity and specificity consistently above 0.7) these eforts have contributed to highlight other fundamental aspects, namely the dificulty to cope 3.3. Calcium score prediction with small and imbalanced datasets. In particular, our Covid-19 study has exacerbated the dificulty to cope Coronary artery disease is the leading cause of death in with small data since the data collection was in progress industrialized countries, despite significant advances in at emergency time during our studies, an issue that af- diagnosis and therapy. In particular, It is now known that fected most of the studies at that time. In fact, during the coronary calcium, indicated with a value called calcium pandemic, it was quite impossible to get balanced and unbiased samples, e.g. the large majority of the admitted 1https://corsa.di.unito.it/. Project funded by Regione Piemonte patients were actually afected by Covid-19, and using Bando INFRA-P2. publicly available data presented many challenges as de- 2https://zenodo.org/record/7501816 3https://zenodo.org/record/5797912 score (CAC), is associated with sub-clinical atheroscle- have shown that DL models, and in particular Deep Neurotic diseases, since its absence is associated with an ral Networks (DNN), largely over-fit site-related noise extremely low and long-lasting probability of cardiovas- when trained on such multi-site datasets, notably due cular events even in subjects at high risk [22]. Calci- to the diference in acquisition protocols, scanner conifcations of the coronary arteries are easily detectable structors, physical properties such as permanent magthrough CT scans, while it is dificult to visually detect netic field [ 25, 26]. This also implies poor generalization them from an X-ray image, even for an expert radiologist: performance on data from new incoming sites, highly It is precisely for this reason that an automatic system limiting the applicability of these models to real-life scefor analyzing X-ray images could be of fundamental im- narios. In order to build more accurate brain age models, portance for the early detection of calcium. Up to our the OpenBHB challenge [27] has been recently released3. knowledge, there is only one study where the presence It is an open-ended challenge, publicly available, which of calcium is inferred from X-rays [23]. In our project we provides one of the largest datasets of healthy brain MRIs. studied the possibility to train properly a CNN able to In this context, together with our partners at Télécom detect the presence of calcium in coronary arteries using Paris and NeuroSpin, CEA, we have developed a novel Chest X-rays as input data instead of CT scan. In partic- contrastive learning loss for regression of brain age from ular, Our model first tries to give an accurate prediction MRI [28]. We validated it on the OpenBHB challenge, of the value of coronary artery calcium (CAC), and based where chronological age must be learned without being on the latter, it determines whether or not a patient has afected by site-related noise. With our method, we obcoronary calcium. This project was developed in collab- tain the best results on the oficial challenge leaderboard. oration with the radiology unit of Città della Salute e della Scienza di Torino (CDSS) hospital in Turin, Who 4.2. Brain perfusion and UnitoBrain collected a dataset of 506 chest X-rays, equally divided between patients with and without coronary calcium.
has become established tool in treatment of ischaemic stroke. During CTP, a series of low-dose scans are ac4. Neuroimaging quired after contrast bolus injection, allowing to compute parametric maps to track perfusion parameters dynamOur research group is also focused on developing innova- ics, e.g. Cerebral Blood Volume (CBV). In our study [29] tive methods and tools for neuroimaging analysis, with a we explored whether a properly trained CNN, based on particular emphasis on Deep Learning (DL) approaches. a U-Net-like structure, can generate informative, paraIn this paper section, we present two subsections show- metric maps such as CBV. The UNITOBrain dataset [30] casing our recent work in the field of neuroimaging. The we created to support the research is publicly available4 ifrst subsection focuses on brain age prediction from MRI, and attracted considerable interest in the area. In the a challenging task that requires robust and accurate mod- end, the agreement between our CNN-based perfusion els capable of generalizing across diferent imaging sites. maps and the state-of-the-art perfusion analysis methods The second subsection explores the use of DL techniques maps based on deconvolution of the data highlights the for the generation of brain perfusion maps from CT im- potential of deep learning methods applied to perfusion ages, aiming to improve the diagnosis and treatment of analysis. Moreover, machine learning methods can reischemic stroke. duce data inputs required to estimate the ischemic core and thus might allow the development of novel perfusion 4.1. Brain age prediction protocols with lower radiation dose. Brain aging involves complex biological processes, such as cortical thinning, that are highly heterogeneous across 5. Omics data for biomarkers isnadmiveidmuaanlsn,esr.ugAgcecsutrinatgeltyhamt opdeeolpinleg dbroainnotagaigneg iant tthhee discovery and computational subject-level is a long-standing goal in neuroscience as it epidemiology could enhance our understanding of age-related diseases such as neurodegenerative disorders. To this end, brain- 5.1. Dataset age predictors linking neuroanatomy to chronological age have been proposed using Deep Learning (DL) [24]. In collaboration with the Italian Istitute for Genomic In order to build accurate biomarker of aging, DL mod- Medicine (IIGM) [31] we collected samples (i.e. stool, els need large-scale neuroimaging dataset for training, cancer tissue and adjacent tissues, plasma) from patients which often involves multi-site studies, partly because of the high cost per patient in each study. Recent works 3https://baobablab.github.io/bhb/ 4https://ieee-dataport.org/open-access/unitobrain at the Clinica S. Rita in Vercelli, Italy. Patients with hered- in preparation). User-defined pipelines of analysis can itary CRC syndromes, with a previous history of CRC, be built by combining the available modules, that impleand with uncompleted or poorly cleaned colonoscopy, ment high-level tasks such as (i) exploratory data analwere excluded from the study. Patients were recruited ysis, (ii) feature selection for biomarker discovery, (iii) at initial diagnosis and had not received any treatment evaluation, for running classification tasks using specific prior to fecal sample collection. Subjects reporting the features and learning algorithms, and (iv) feature extracuse of antibiotics during the 6 months prior to the sample tion, for building models based on human-interpretable collection were excluded from the study. On the basis features in the form of logical rules. The basic idea is of colonoscopy results, recruited subjects were classified to provide a tool for automated machine learning usable into three categories: 1) healthy subjects: individuals by researchers without computational skills, in order to with colonoscopy negative for tumor, adenomas, and help them in the functional interpretation of results from other diseases; 2) adenoma patients: individuals with a clinical/biological point of view. colorectal adenoma/s; and 3) CRC patients: individuals with newly diagnosed CRC. A total of 93 subjects were 5.3. Computational Epidemiology initially recruited, and the 80 that passed quality control are divided into 29 CRC patients, 27 adenomas, and 24 Computational epidemiology exploits Artificial Intellicontrols. On the samples collected, the shotgun and small gence and Simulation to successfully support epideminoncoding RNA sequencing were performed. ologists, healthcare professionals, and decision-makers
The main results are reported firstly in [ 32] where we to understand and control the spatio-temporal spread of identified a specific signature composed of profiles of infectious diseases. In particular, during the first phase of human small non-coding RNAs, microbial sRNAs, and the COVID-19 outbreak, the qBio group was involved, in microbial DNAs was able to accurately classify the three collaboration with the Department of Medical Sciences of categories of subjects with a high level of performance. the University of Turin, to support decision-makers of the
This evidence was confirmed across multiple cohorts. Italian Piedmont region for evaluating the impact of difIndeed, we assessed the CRC-associated gut microbiome ferent implementations of the infection control measures and its ability to distinguish newly diagnosed CRC pa- [? ] (e.g., non-pharmaceutical interventions, surveillance tients from tumor-free controls. Our study [33] was per- methods, and screening tests) by exploiting the general formed across nine multiple datasets and a combined modelling framework GreatMod [34]. Most recently analysis based on Random Forest based machine learn- qBio’s members are developing an integrated smart dashing approach. The identification of reproducible micro- board for providing early warnings, monitoring and forebial biomarkers for CRC may enable the design of non- casting tools to public health response agencies and local invasive diagnostic tools. healthcare services for anticipating medical needs. 5.2. Multi-omics data integration
huge amounts of diferent omics, each one providing complementary perspectives of the biological system under study. Omics data analysis is challenging because of its high dimensionality, noisiness, and error-proneness.
Moreover, most diseases and phenomena afect complex molecular pathways where multiple omics interact with each other, and multi-omics integration can provide a more comprehensive understanding of the biological system under study. Machine learning provides the methodologies to eficiently tackle all these challenges. However, it is dificult to choose the most appropriate algorithm for the data collected, and for the subsequent data integration. The main challenges of such a task are linked to the complexity, heterogeneity, dynamics, uncertainty and high dimensionality, as well as to the right methodologies to analyze and integrate such data.
In this contest, the q-Bio group developed a modular framework for multi-omics integration, called FeatSEE (Feature Selection, Evaluation, and Explanation), paper nosis and prediction of treatment response., https: [21] H. A. H. Chaudhry, R. Renzulli, D. Perlo, F. San//www.oncobiome.eu/, 2019. tinelli, S. Tibaldi, C. Cristiano, M. Grosso, A. Fian[9] R. S. Gonzalez, Updates and challenges in gastroin- drotti, M. Lucenteforte, D. Cavagnino, Lung nodtestinal pathology, Surgical Pathology Clinics 13 ules segmentation with deephealth toolkit, in: (2020) ix. ICIAP, 2022. [10] B. Denis, C. Peters, C. Chapelain, I. Kleinclaus, [22] L. A. anf Alessandro Mazzei, Coronary artery calA. Fricker, R. Wild, B. Auge, I. Gendre, P. Perrin, cium scoring: Its practicality and clinical utility in D. Chatelain, et al., Diagnostic accuracy of commu- primary care, in: Progress in Artificial Intelligence nity pathologists in the interpretation of colorectal - 17th Portuguese Conference on Artificial Intellipolyps, European journal of gastroenterology & gence, EPIA 2015, Coimbra, Portugal, September hepatology 21 (2009) 1153–1160. 8-11, 2015. Proceedings, 2015. [11] D. Perlo, E. Tartaglione, L. Bertero, P. Cassoni, [23] H. S. Peter Kamel, Paul Yi, C. T. Lin, Prediction of M. Grangetto, Dysplasia grading of colorectal coronary artery calcium and cardiovascular risk on polyps through convolutional neural network anal- chest radiographs using deep learning, Radiology: ysis of whole slide images, in: Proceedings of Int. Cardiothoracic Imaging (2021).
Conf. on Medical Imaging and Computer-Aided Di- [24] H. Peng, et al., Accurate brain age prediction with agnosis, 2022, pp. 325–334. lightweight deep neural networks, MedIA (2021). [12] C. A. Barbano, D. Perlo, E. Tartaglione, A. Fiandrotti, [25] C. Wachinger, et al., Detect and correct bias in L. Bertero, P. Cassoni, M. Grangetto, Unitopatho, multi-site neuroimaging datasets, MedIA (2021). a labeled histopathological dataset for colorectal [26] B. Glocker, et al., Machine learning with multi-site polyps classification and adenoma dysplasia grad- imaging data: An empirical study on the impact of ing, in: IEEE ICIP, 2021. scanner efects, in: MedNeurIPS Workshop, 2019. [13] L. Bertero, C. Barbano, et. al., Unitopatho, 2021. [27] B. Dufumier, et al., Openbhb: a large-scale multidoi:1 0 . 2 1 2 2 7 / 9 f s v - t m 2 5 . site brain mri data-set for age prediction and debi[14] C. A. Barbano, Eidoslab/torchstain: v1.2.0-stable, asing, NeuroImage (2022).
2022. URL: https://doi.org/10.5281/zenodo.6979540. [28] C. A. Barbano, B. Dufumier, E. Duchesnay, doi:1 0 . 5 2 8 1 / z e n o d o . 6 9 7 9 5 4 0 . M. Grangetto, P. Gori, Contrastive learning for [15] Z. Cai, R. C. Poulos, J. Liu, Q. Zhong, Ma- regression in multi-site brain age prediction, 2023. chine learning for multi-omics data in- [29] U. A. Gava, F. D’Agata, E. Tartaglione, R. Renzulli, tegration in cancer, iScience 25 (2022) M. Grangetto, F. Bertolino, A. Santonocito, E. Ben103798. URL: https://www.sciencedirect. nink, G. Vaudano, A. Boghi, M. Bergui, Neural com/science/article/pii/S2589004222000682. network-derived perfusion maps: a model-free apdoi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . i s c i . 2 0 2 2 . 1 0 3 7 9 8 . proach to computed tomography perfusion in pa[16] E. Tartaglione, C. A. Barbano, C. Berzovini, M. Ca- tients with acute ischemic stroke, Frontiers in Neulandri, M. Grangetto, Unveiling covid-19 from chest roinformatics (2023). x-ray with deep learning: a hurdles race with small [30] D. Perlo, E. Tartaglione, U. Gava, F. D’Agata, E. Bendata, International Journal of Environmental Re- ninck, M. Bergui, Unitobrain dataset: A brain persearch and Public Health 17 (2020) 6933. fusion dataset, in: ICIAP, 2022. [17] C. A. Barbano, E. Tartaglione, C. Berzovini, M. Ca- [31] IIGM, Italian Istitute for Genomic Medicine, https: landri, M. Grangetto, A two-step radiologist-like //www.iigm.it/site/, 2021. approach for covid-19 computer-aided diagnosis [32] S. Tarallo, G. Ferrero, G. Gallo, et. al., Altered fecal from chest x-ray images, in: ICIAP, 2022. small rna profiles in colorectal cancer reflect gut mi[18] E. Tartaglione, C. A. Barbano, M. Grangetto, End: crobiome composition in stool samples., mSystems Entangling and disentangling deep representations (2019). for bias correction, in: IEEE Conference on Com- [33] A. Thomas, P. Manghi, F. Asnicar, et. al, Metageputer Vision and Pattern Recognition (CVPR), 2021, nomic analysis of colorectal cancer datasets identi2021. ifes cross-cohort microbial diagnostic signatures [19] R. L. Siegel, K. D. Miller, H. E. Fuchs, A. Jemal, Can- and a link with choline degradation, Nature cer statistics, 2021, CA: A Cancer Journal for Clini- Medicine (2019).
cians 71 (2021) 7–33. doi:1 0 . 3 3 2 2 / c a a c . 2 1 6 5 4 . [34] S. Pernice, P. Castagno, L. Marcotulli, M. Maule, [20] H. A. H. Chaudhry, R. Renzulli, D. Perlo, F. San- L. Richiardi, G. Moirano, M. Sereno, F. Cordero, tinelli, S. Tibaldi, C. Cristiano, M. Grosso, M. Beccuti, Impacts of reopening strategies for G. Limerutti, A. Fiandrotti, M. Grangetto, et al., Uni- covid-19 epidemic: a modeling study in piedmont tochest: A lung image dataset for segmentation of region, BMC Infectious Diseases (2020). cancerous nodules on ct scans, in: ICIAP, 2022.