Dmytro Fishman
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Dmytro Fishman has a total of 78 co-authors. Here are the top co-authors by number of shared publications:
- Leopold Parts
Wellcome Sanger Institute, Wellcome Genome Campus, Hinxton, UK
5 shared publications - Mohammed Abduljabbar Ali
Department of Computer Science, University of Tartu, Tartu, Estonia
5 shared publications - Tõnis Laasfeld
Institute of Chemistry, University of Tartu
4 shared publications - Ago Rinken
Institute of Chemistry, University of Tartu, Tartu, Estonia
3 shared publications - Hedi Peterson
Institute of Computer Science, University of Tartu
3 shared publications - Jane Torp
Institute of Structural Biology, University of Bonn, Bonn, Germany
3 shared publications
Latest Publications
- Open Access
- 27 November 2025
- Published by Springer Nature in Communications Medicine
- Vol. 5 (1)
A clinically validated AI framework for kidney cancer detection and characterization
Background Renal cell carcinoma is one of the most common cancers of the urinary tract and is usually diagnosed by interpreting contrast-enhanced computed tomography scans. Rising demand for radiological services, combined with a shortage of radiologists, makes timely and accurate diagnosis increasingly challenging. Automated approaches may help radiologists improve efficiency and accuracy. Methods We developed BMVision, a deep learning-based tool for detecting and characterizing kidney cancer. The tool integrates with a web-based viewer designed to provide an intuitive interface for radiologists. Its performance was evaluated in a two-stage retrospective reader study. Six radiologists independently reviewed 200 scans across both AI-assisted and unaided workflows, allowing comparison of diagnostic performance and workflow efficiency with and without support from the tool. Statistical analysis compared AI-aided and unaided workflows across predefined clinical criteria, including diagnostic sensitivity, lesion measurement, reporting efficiency, and inter-radiologist agreement, using non-parametric tests and bootstrapping. Results Here we show that BMVision reduces radiologists’ reporting time by an average of 33%, up to 52%. The tool provides structured auto-generated reports, minimizing the need for manual dictation or typing. In addition, BMVision improves sensitivity for detecting benign renal lesions (from 79.9 to 86.3%) and leads to a significant increase in inter-radiologist agreement. Conclusions To the best of our knowledge, BMVision is the first clinically validated commercial artificial intelligence tool for kidney cancer detection and characterization. By improving diagnostic accuracy and reporting efficiency, it has the potential to enhance patient care and help radiologists meet the growing demand for high-quality cancer diagnostics.
- 27 November 2025
- Published by Springer Nature in Communications Medicine
- Vol. 5 (1)
Open AccessA clinically validated AI framework for kidney cancer detection and characterization
Background Renal cell carcinoma is one of the most common cancers of the urinary tract and is usually diagnosed by interpreting contrast-enhanced computed tomography scans. Rising demand for radiological services, combined with a shortage of radiologists, makes timely and accurate diagnosis increasingly challenging. Automated approaches may help radiologists improve efficiency and accuracy. Methods We developed BMVision, a deep learning-based tool for detecting and characterizing kidney cancer. The tool integrates with a web-based viewer designed to provide an intuitive interface for radiologists. Its performance was evaluated in a two-stage retrospective reader study. Six radiologists independently reviewed 200 scans across both AI-assisted and unaided workflows, allowing comparison of diagnostic performance and workflow efficiency with and without support from the tool. Statistical analysis compared AI-aided and unaided workflows across predefined clinical criteria, including diagnostic sensitivity, lesion measurement, reporting efficiency, and inter-radiologist agreement, using non-parametric tests and bootstrapping. Results Here we show that BMVision reduces radiologists’ reporting time by an average of 33%, up to 52%. The tool provides structured auto-generated reports, minimizing the need for manual dictation or typing. In addition, BMVision improves sensitivity for detecting benign renal lesions (from 79.9 to 86.3%) and leads to a significant increase in inter-radiologist agreement. Conclusions To the best of our knowledge, BMVision is the first clinically validated commercial artificial intelligence tool for kidney cancer detection and characterization. By improving diagnostic accuracy and reporting efficiency, it has the potential to enhance patient care and help radiologists meet the growing demand for high-quality cancer diagnostics.
- Open Access
- 28 September 2025
- Published by Springer Nature
- p. 87-96
Interpreting CT-Scans with CLIP: An Explorative Study of Attribution Methods for 3D Vision-Language Models
- 28 September 2025
- Published by Springer Nature
- p. 87-96
Open AccessInterpreting CT-Scans with CLIP: An Explorative Study of Attribution Methods for 3D Vision-Language Models
- Open Access
- 27 June 2025
- Published by Elsevier in European Journal of Pharmaceutical Sciences
- Vol. 212
FiBar: A tool for analyzing fiber diameters in complex drug delivery systems from scanning electron microscopy images
- 27 June 2025
- Published by Elsevier in European Journal of Pharmaceutical Sciences
- Vol. 212
Open AccessFiBar: A tool for analyzing fiber diameters in complex drug delivery systems from scanning electron microscopy images
- 12 May 2025
- Published by Springer Nature
- p. 222-238
Going Beyond U-Net: Assessing Vision Transformers for Semantic Segmentation in Microscopy Image Analysis
- 12 May 2025
- Published by Springer Nature
- p. 222-238
Going Beyond U-Net: Assessing Vision Transformers for Semantic Segmentation in Microscopy Image Analysis
- Open Access
Going Beyond U-Net: Assessing Vision Transformers for Semantic Segmentation in Microscopy Image Analysis
Segmentation is a crucial step in microscopy image analysis. Numerous approaches have been developed over the past years, ranging from classical segmentation algorithms to advanced deep learning models. While U-Net remains one of the most popular and well-established models for biomedical segmentation tasks, recently developed transformer-based models promise to enhance the segmentation process of microscopy images. In this work, we assess the efficacy of transformers, including UNETR, the Segment Anything Model, and Swin-UPerNet, and compare them with the well-established U-Net model across various image modalities such as electron microscopy, brightfield, histopathology, and phase-contrast. Our evaluation identifies several limitations in the original Swin Transformer model, which we address through architectural modifications to optimise its performance. The results demonstrate that these modifications improve segmentation performance compared to the classical U-Net model and the unmodified Swin-UPerNet. This comparative analysis highlights the promise of transformer models for advancing biomedical image segmentation. It demonstrates that their efficiency and applicability can be improved with careful modifications, facilitating their future use in microscopy image analysis tools.
Open AccessGoing Beyond U-Net: Assessing Vision Transformers for Semantic Segmentation in Microscopy Image Analysis
Segmentation is a crucial step in microscopy image analysis. Numerous approaches have been developed over the past years, ranging from classical segmentation algorithms to advanced deep learning models. While U-Net remains one of the most popular and well-established models for biomedical segmentation tasks, recently developed transformer-based models promise to enhance the segmentation process of microscopy images. In this work, we assess the efficacy of transformers, including UNETR, the Segment Anything Model, and Swin-UPerNet, and compare them with the well-established U-Net model across various image modalities such as electron microscopy, brightfield, histopathology, and phase-contrast. Our evaluation identifies several limitations in the original Swin Transformer model, which we address through architectural modifications to optimise its performance. The results demonstrate that these modifications improve segmentation performance compared to the classical U-Net model and the unmodified Swin-UPerNet. This comparative analysis highlights the promise of transformer models for advancing biomedical image segmentation. It demonstrates that their efficiency and applicability can be improved with careful modifications, facilitating their future use in microscopy image analysis tools.
- 10 July 2024
- Published by Springer Nature
- p. 39-59
COIN: Counterfactual Inpainting for Weakly Supervised Semantic Segmentation for Medical Images
- 10 July 2024
- Published by Springer Nature
- p. 39-59
COIN: Counterfactual Inpainting for Weakly Supervised Semantic Segmentation for Medical Images
- 14 October 2023
- Published by Springer Nature
- p. 147-155
Metadata Improves Segmentation Through Multitasking Elicitation
- 14 October 2023
- Published by Springer Nature
- p. 147-155
Metadata Improves Segmentation Through Multitasking Elicitation
- Open Access
Untargeted metabolomics profiling in pediatric patients and adult populations indicates a connection between lipid imbalance and epilepsy
Introduction: Epilepsy is a common central nervous system disorder characterized by abnormal brain electrical activity. We aimed to compare the metabolic profiles of plasma from patients with epilepsy across different etiologies, seizure frequency, seizure type, and patient age to try to identify common disrupted pathways. Material and methods: We used data from three separate cohorts. The first cohort (PED-C) consisted of 31 pediatric patients with suspicion of a genetic disorder with unclear etiology; the second cohort (AD-C) consisted of 250 adults from the Estonian Biobank (EstBB), and the third cohort consisted of 583 adults ≥ 69 years of age from the EstBB (ELD-C). We compared untargeted metabolomics and lipidomics data between individuals with and without epilepsy in each cohort. Results: In the PED-C, significant alterations (p-value <0.05) were detected in sixteen different glycerophosphatidylcholines (GPC), dimethylglycine and eicosanedioate (C20-DC). In the AD-C, nine significantly altered metabolites were found, mainly triacylglycerides (TAG), which are also precursors in the GPC synthesis pathway. In the ELD-C, significant changes in twenty metabolites including multiple TAGs were observed in the metabolic profile of participants with previously diagnosed epilepsy. Pathway analysis revealed that among the metabolites that differ significantly between epilepsy-positive and epilepsy-negative patients in the PED-C, the lipid superpathway (p = 3.2*10-4) and phosphatidylcholine (p = 9.3*10-8) and lysophospholipid (p = 5.9*10-3) subpathways are statistically overrepresented. Analogously, in the AD-C, the triacylglyceride subclass turned out to be statistically overrepresented (p = 8.5*10-5) with the lipid superpathway (p = 1.4*10-2). The presented p-values are FDR-corrected. Conclusion: Our results suggest that cell membrane fluidity may have a significant role in the mechanism of epilepsy, and changes in lipid balance may indicate epilepsy. However, further studies are needed to evaluate whether untargeted metabolomics analysis could prove helpful in diagnosing epilepsy earlier.
Open AccessUntargeted metabolomics profiling in pediatric patients and adult populations indicates a connection between lipid imbalance and epilepsy
Introduction: Epilepsy is a common central nervous system disorder characterized by abnormal brain electrical activity. We aimed to compare the metabolic profiles of plasma from patients with epilepsy across different etiologies, seizure frequency, seizure type, and patient age to try to identify common disrupted pathways. Material and methods: We used data from three separate cohorts. The first cohort (PED-C) consisted of 31 pediatric patients with suspicion of a genetic disorder with unclear etiology; the second cohort (AD-C) consisted of 250 adults from the Estonian Biobank (EstBB), and the third cohort consisted of 583 adults ≥ 69 years of age from the EstBB (ELD-C). We compared untargeted metabolomics and lipidomics data between individuals with and without epilepsy in each cohort. Results: In the PED-C, significant alterations (p-value <0.05) were detected in sixteen different glycerophosphatidylcholines (GPC), dimethylglycine and eicosanedioate (C20-DC). In the AD-C, nine significantly altered metabolites were found, mainly triacylglycerides (TAG), which are also precursors in the GPC synthesis pathway. In the ELD-C, significant changes in twenty metabolites including multiple TAGs were observed in the metabolic profile of participants with previously diagnosed epilepsy. Pathway analysis revealed that among the metabolites that differ significantly between epilepsy-positive and epilepsy-negative patients in the PED-C, the lipid superpathway (p = 3.2*10-4) and phosphatidylcholine (p = 9.3*10-8) and lysophospholipid (p = 5.9*10-3) subpathways are statistically overrepresented. Analogously, in the AD-C, the triacylglyceride subclass turned out to be statistically overrepresented (p = 8.5*10-5) with the lipid superpathway (p = 1.4*10-2). The presented p-values are FDR-corrected. Conclusion: Our results suggest that cell membrane fluidity may have a significant role in the mechanism of epilepsy, and changes in lipid balance may indicate epilepsy. However, further studies are needed to evaluate whether untargeted metabolomics analysis could prove helpful in diagnosing epilepsy earlier.
- Open Access
- 6 July 2022
- Published by Springer Nature in Scientific Reports
- Vol. 12 (1) , 1-13
ArtSeg—Artifact segmentation and removal in brightfield cell microscopy images without manual pixel-level annotations
Brightfield cell microscopy is a foundational tool in life sciences. The acquired images are prone to contain visual artifacts that hinder downstream analysis, and automatically removing them is therefore of great practical interest. Deep convolutional neural networks are state-of-the-art for image segmentation, but require pixel-level annotations, which are time-consuming to produce. Here, we propose ScoreCAM-U-Net, a pipeline to segment artifactual regions in brightfield images with limited user input. The model is trained using only image-level labels, so the process is faster by orders of magnitude compared to pixel-level annotation, but without substantially sacrificing the segmentation performance. We confirm that artifacts indeed exist with different shapes and sizes in three different brightfield microscopy image datasets, and distort downstream analyses such as nuclei segmentation, morphometry and fluorescence intensity quantification. We then demonstrate that our automated artifact removal ameliorates this problem. Such rapid cleaning of acquired images using the power of deep learning models is likely to become a standard step for all large scale microscopy experiments.
- 6 July 2022
- Published by Springer Nature in Scientific Reports
- Vol. 12 (1) , 1-13
Open AccessArtSeg—Artifact segmentation and removal in brightfield cell microscopy images without manual pixel-level annotations
Brightfield cell microscopy is a foundational tool in life sciences. The acquired images are prone to contain visual artifacts that hinder downstream analysis, and automatically removing them is therefore of great practical interest. Deep convolutional neural networks are state-of-the-art for image segmentation, but require pixel-level annotations, which are time-consuming to produce. Here, we propose ScoreCAM-U-Net, a pipeline to segment artifactual regions in brightfield images with limited user input. The model is trained using only image-level labels, so the process is faster by orders of magnitude compared to pixel-level annotation, but without substantially sacrificing the segmentation performance. We confirm that artifacts indeed exist with different shapes and sizes in three different brightfield microscopy image datasets, and distort downstream analyses such as nuclei segmentation, morphometry and fluorescence intensity quantification. We then demonstrate that our automated artifact removal ameliorates this problem. Such rapid cleaning of acquired images using the power of deep learning models is likely to become a standard step for all large scale microscopy experiments.