Andrea Espis
Research fellow
Area of interest:
My research focuses on the development of artificial intelligence models for the integration of heterogeneous biomedical data, including proteomic, metagenomic, genomic and metabolomic profiles, hormonal profiles, clinical data, and dietary and lifestyle information.
Building on my experience in self-supervised, multi-instance and multimodal deep learning for medical imaging, digital pathology and genomics, I aim to develop robust and reproducible methods that uncover patterns shared across omic layers and over time, and that help formulate and test causal hypotheses.
Most significant publications:
2026
Espis, Andrea; Marzi, Chiara; Diciotti, Stefano
NeuroBooster: a domain-informed self-supervised learning paradigm tailored for brain MRI analysis Journal Article
In: IEEE J. Biomed. Health Inform., vol. PP, pp. 1–14, 2026.
@article{Espis2026-cr,
title = {NeuroBooster: a domain-informed self-supervised learning
paradigm tailored for brain MRI analysis},
author = {Andrea Espis and Chiara Marzi and Stefano Diciotti},
year = {2026},
date = {2026-06-01},
journal = {IEEE J. Biomed. Health Inform.},
volume = {PP},
pages = {1–14},
publisher = {Institute of Electrical and Electronics Engineers (IEEE)},
abstract = {Self-supervised learning (SSL) has demonstrated its potential to
reduce reliance on labeled data and improve generalization to
independent test sets compared to traditional supervised
learning (SL). However, current SSL paradigms are predominantly
designed as domain-agnostic and are often evaluated on benchmark
datasets like ImageNet. As already shown in other studies, SSL
paradigms tailored for the medical domain can outperform
domain-agnostic SSL paradigms. This study introduces
NeuroBooster, a paradigm tailored for brain MRI analysis. Its
core idea is to leverage automated feature extraction tools to
generate anatomical features, which are then used as targets in
a regression pretext task. The paradigms were evaluated and
compared on a regression task, i.e., age prediction, and a
classification task, i.e., the discrimination between patients
with Alzheimer's disease and cognitively normal subjects. Each
pre-training and fine tuning experiment was repeated across 30
randomized seeds to estimate the uncertainty, while enabling
full reproducibility. Our results revealed NeuroBooster's
superiority across various scenarios. Notably, with only 1% of
labeled data available, NeuroBooster achieved an average mean
absolute error of 5.67 years smaller than SL. These findings
demonstrate NeuroBooster's efficacy and highlight its potential
within brain MRI analysis, suggesting a promising direction for
developing specialized SSL paradigms in this domain.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Self-supervised learning (SSL) has demonstrated its potential to
reduce reliance on labeled data and improve generalization to
independent test sets compared to traditional supervised
learning (SL). However, current SSL paradigms are predominantly
designed as domain-agnostic and are often evaluated on benchmark
datasets like ImageNet. As already shown in other studies, SSL
paradigms tailored for the medical domain can outperform
domain-agnostic SSL paradigms. This study introduces
NeuroBooster, a paradigm tailored for brain MRI analysis. Its
core idea is to leverage automated feature extraction tools to
generate anatomical features, which are then used as targets in
a regression pretext task. The paradigms were evaluated and
compared on a regression task, i.e., age prediction, and a
classification task, i.e., the discrimination between patients
with Alzheimer's disease and cognitively normal subjects. Each
pre-training and fine tuning experiment was repeated across 30
randomized seeds to estimate the uncertainty, while enabling
full reproducibility. Our results revealed NeuroBooster's
superiority across various scenarios. Notably, with only 1% of
labeled data available, NeuroBooster achieved an average mean
absolute error of 5.67 years smaller than SL. These findings
demonstrate NeuroBooster's efficacy and highlight its potential
within brain MRI analysis, suggesting a promising direction for
developing specialized SSL paradigms in this domain.
reduce reliance on labeled data and improve generalization to
independent test sets compared to traditional supervised
learning (SL). However, current SSL paradigms are predominantly
designed as domain-agnostic and are often evaluated on benchmark
datasets like ImageNet. As already shown in other studies, SSL
paradigms tailored for the medical domain can outperform
domain-agnostic SSL paradigms. This study introduces
NeuroBooster, a paradigm tailored for brain MRI analysis. Its
core idea is to leverage automated feature extraction tools to
generate anatomical features, which are then used as targets in
a regression pretext task. The paradigms were evaluated and
compared on a regression task, i.e., age prediction, and a
classification task, i.e., the discrimination between patients
with Alzheimer's disease and cognitively normal subjects. Each
pre-training and fine tuning experiment was repeated across 30
randomized seeds to estimate the uncertainty, while enabling
full reproducibility. Our results revealed NeuroBooster's
superiority across various scenarios. Notably, with only 1% of
labeled data available, NeuroBooster achieved an average mean
absolute error of 5.67 years smaller than SL. These findings
demonstrate NeuroBooster's efficacy and highlight its potential
within brain MRI analysis, suggesting a promising direction for
developing specialized SSL paradigms in this domain.
Roncarati, Roberta; Fontana, Beatrice; Pace, Ilaria; Durante, Giorgio; Conci, Nicole; Soru, Alessia; Gallerani, Giulia; Riefolo, Mattia; Salamon, Irene; Laprovitera, Noemi; Broseghini, Elisabetta; Naddeo, Maria; Espis, Andrea; Sales, Gabriele; Diciotti, Stefano; D'Errico, Antonia; Rihawi, Karim; Ardizzoni, Andrea; Ferracin, Manuela
Cancer of unknown primary genomic profiling from cell-free DNA provides insights into CUP biology and vulnerabilities Journal Article
In: Front. Med. (Lausanne), vol. 13, no. 1777582, pp. 1777582, 2026.
@article{Roncarati2026-yb,
title = {Cancer of unknown primary genomic profiling from cell-free DNA
provides insights into CUP biology and vulnerabilities},
author = {Roberta Roncarati and Beatrice Fontana and Ilaria Pace and Giorgio Durante and Nicole Conci and Alessia Soru and Giulia Gallerani and Mattia Riefolo and Irene Salamon and Noemi Laprovitera and Elisabetta Broseghini and Maria Naddeo and Andrea Espis and Gabriele Sales and Stefano Diciotti and Antonia D'Errico and Karim Rihawi and Andrea Ardizzoni and Manuela Ferracin},
year = {2026},
date = {2026-06-01},
journal = {Front. Med. (Lausanne)},
volume = {13},
number = {1777582},
pages = {1777582},
publisher = {Frontiers Media SA},
abstract = {Introduction: Cancer of unknown primary (CUP) is a metastatic
malignancy with no identifiable site of origin, accounting for
2-5% of cancer diagnoses. Its marked heterogeneity and the
limited availability of tumor tissue pose major challenges to
genomic profiling. Methods: In this retrospective observational
study, we applied a 92-gene CUP-specific targeted sequencing
panel to liquid biopsy samples from 39 CUP patients, analyzing
circulating cell-free DNA (ccfDNA) together with paired germline
DNA (gDNA), when available. Somatic, germline, and CHIP-related
variants were classified using predefined variant allele
frequency (VAF) thresholds and paired ccfDNA/gDNA comparison.
Results: We identified somatic mutations in 78 of 92 genes and
44 clinically relevant variants (Tier I-III), most frequently
affecting NF1, KRAS, ARID1A, and PIK3CA. Mutated genes were
primarily involved in cell-cycle regulation, receptor tyrosine
kinase signaling, and NOTCH pathways. Recurrent genetic
alterations were detected in 15 genes, including canonical
hotspot mutations in TP53, KRAS, and PIK3CA, 10 of which were
shared by three or more patients. Actionable mutations, as
defined by current clinical annotation guidelines, were
identified in 13 genes, supporting the potential role of
molecularly guided therapies in CUP. Likely CHIP-associated
variants, defined as mutations in CHIP-associated genes with VAF
$geq$ 2% in PBMC-derived genomic DNA, were detected in 17
patients. In addition, pathogenic germline variants in MITF,
NTRK1, and BAP1 were identified in three patients; notably, the
NTRK1 germline variant was accompanied by an independent somatic
mutation in the same gene. Discussion: Overall, these findings
support liquid biopsy as a valuable approach for molecular
profiling of CUP and highlight the critical importance of paired
ccfDNA/gDNA analysis, including CHIP assessment, to accurately
distinguish somatic, germline, and hematopoiesis-related
variants.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Introduction: Cancer of unknown primary (CUP) is a metastatic
malignancy with no identifiable site of origin, accounting for
2-5% of cancer diagnoses. Its marked heterogeneity and the
limited availability of tumor tissue pose major challenges to
genomic profiling. Methods: In this retrospective observational
study, we applied a 92-gene CUP-specific targeted sequencing
panel to liquid biopsy samples from 39 CUP patients, analyzing
circulating cell-free DNA (ccfDNA) together with paired germline
DNA (gDNA), when available. Somatic, germline, and CHIP-related
variants were classified using predefined variant allele
frequency (VAF) thresholds and paired ccfDNA/gDNA comparison.
Results: We identified somatic mutations in 78 of 92 genes and
44 clinically relevant variants (Tier I-III), most frequently
affecting NF1, KRAS, ARID1A, and PIK3CA. Mutated genes were
primarily involved in cell-cycle regulation, receptor tyrosine
kinase signaling, and NOTCH pathways. Recurrent genetic
alterations were detected in 15 genes, including canonical
hotspot mutations in TP53, KRAS, and PIK3CA, 10 of which were
shared by three or more patients. Actionable mutations, as
defined by current clinical annotation guidelines, were
identified in 13 genes, supporting the potential role of
molecularly guided therapies in CUP. Likely CHIP-associated
variants, defined as mutations in CHIP-associated genes with VAF
$geq$ 2% in PBMC-derived genomic DNA, were detected in 17
patients. In addition, pathogenic germline variants in MITF,
NTRK1, and BAP1 were identified in three patients; notably, the
NTRK1 germline variant was accompanied by an independent somatic
mutation in the same gene. Discussion: Overall, these findings
support liquid biopsy as a valuable approach for molecular
profiling of CUP and highlight the critical importance of paired
ccfDNA/gDNA analysis, including CHIP assessment, to accurately
distinguish somatic, germline, and hematopoiesis-related
variants.
malignancy with no identifiable site of origin, accounting for
2-5% of cancer diagnoses. Its marked heterogeneity and the
limited availability of tumor tissue pose major challenges to
genomic profiling. Methods: In this retrospective observational
study, we applied a 92-gene CUP-specific targeted sequencing
panel to liquid biopsy samples from 39 CUP patients, analyzing
circulating cell-free DNA (ccfDNA) together with paired germline
DNA (gDNA), when available. Somatic, germline, and CHIP-related
variants were classified using predefined variant allele
frequency (VAF) thresholds and paired ccfDNA/gDNA comparison.
Results: We identified somatic mutations in 78 of 92 genes and
44 clinically relevant variants (Tier I-III), most frequently
affecting NF1, KRAS, ARID1A, and PIK3CA. Mutated genes were
primarily involved in cell-cycle regulation, receptor tyrosine
kinase signaling, and NOTCH pathways. Recurrent genetic
alterations were detected in 15 genes, including canonical
hotspot mutations in TP53, KRAS, and PIK3CA, 10 of which were
shared by three or more patients. Actionable mutations, as
defined by current clinical annotation guidelines, were
identified in 13 genes, supporting the potential role of
molecularly guided therapies in CUP. Likely CHIP-associated
variants, defined as mutations in CHIP-associated genes with VAF
$geq$ 2% in PBMC-derived genomic DNA, were detected in 17
patients. In addition, pathogenic germline variants in MITF,
NTRK1, and BAP1 were identified in three patients; notably, the
NTRK1 germline variant was accompanied by an independent somatic
mutation in the same gene. Discussion: Overall, these findings
support liquid biopsy as a valuable approach for molecular
profiling of CUP and highlight the critical importance of paired
ccfDNA/gDNA analysis, including CHIP assessment, to accurately
distinguish somatic, germline, and hematopoiesis-related
variants.
2025
Espis, Andrea; Marzi, Chiara; Diciotti, Stefano
Comparative analysis of supervised and self-supervised learning with small and imbalanced medical imaging datasets Journal Article
In: Sci. Rep., vol. 15, no. 1, pp. 32345, 2025.
@article{Espis2025-xw,
title = {Comparative analysis of supervised and self-supervised learning
with small and imbalanced medical imaging datasets},
author = {Andrea Espis and Chiara Marzi and Stefano Diciotti},
year = {2025},
date = {2025-09-01},
journal = {Sci. Rep.},
volume = {15},
number = {1},
pages = {32345},
publisher = {Springer Science and Business Media LLC},
abstract = {Self-supervised learning (SSL) in computer vision has shown its
potential to reduce reliance on labeled data. However, most
studies focused on balanced, large, broad-domain datasets like
ImageNet, whereas, in real-world medical applications, dataset
size is typically limited. This study compares the performance
of SSL versus supervised learning (SL) on small, imbalanced
medical imaging datasets. We experimented with four binary
classification tasks: age prediction and diagnosis of
Alzheimer's disease from brain magnetic resonance imaging scans,
pneumonia from chest radiograms, and retinal diseases associated
with choroidal neovascularization from optical coherence
tomography with a mean size of training sets of 843 images, 771
images, 1,214 images, and 33,484 images, respectively. We tested
various combinations of label availability and class frequency
distribution, repeating the training with different random seeds
to assess result uncertainty. In most experiments involving
small training sets, SL outperformed the selected SSL paradigms,
even when a limited portion of labeled data was available. Our
findings highlight the importance of carefully selecting
learning paradigms based on specific application requirements,
which are influenced by factors such as training set size, label
availability, and class frequency distribution.},
keywords = {},
pubstate = {published},
tppubtype = {article}
}
Self-supervised learning (SSL) in computer vision has shown its
potential to reduce reliance on labeled data. However, most
studies focused on balanced, large, broad-domain datasets like
ImageNet, whereas, in real-world medical applications, dataset
size is typically limited. This study compares the performance
of SSL versus supervised learning (SL) on small, imbalanced
medical imaging datasets. We experimented with four binary
classification tasks: age prediction and diagnosis of
Alzheimer's disease from brain magnetic resonance imaging scans,
pneumonia from chest radiograms, and retinal diseases associated
with choroidal neovascularization from optical coherence
tomography with a mean size of training sets of 843 images, 771
images, 1,214 images, and 33,484 images, respectively. We tested
various combinations of label availability and class frequency
distribution, repeating the training with different random seeds
to assess result uncertainty. In most experiments involving
small training sets, SL outperformed the selected SSL paradigms,
even when a limited portion of labeled data was available. Our
findings highlight the importance of carefully selecting
learning paradigms based on specific application requirements,
which are influenced by factors such as training set size, label
availability, and class frequency distribution.
potential to reduce reliance on labeled data. However, most
studies focused on balanced, large, broad-domain datasets like
ImageNet, whereas, in real-world medical applications, dataset
size is typically limited. This study compares the performance
of SSL versus supervised learning (SL) on small, imbalanced
medical imaging datasets. We experimented with four binary
classification tasks: age prediction and diagnosis of
Alzheimer's disease from brain magnetic resonance imaging scans,
pneumonia from chest radiograms, and retinal diseases associated
with choroidal neovascularization from optical coherence
tomography with a mean size of training sets of 843 images, 771
images, 1,214 images, and 33,484 images, respectively. We tested
various combinations of label availability and class frequency
distribution, repeating the training with different random seeds
to assess result uncertainty. In most experiments involving
small training sets, SL outperformed the selected SSL paradigms,
even when a limited portion of labeled data was available. Our
findings highlight the importance of carefully selecting
learning paradigms based on specific application requirements,
which are influenced by factors such as training set size, label
availability, and class frequency distribution.
- Monserrato
070 6754593