Ashish Singh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ashish Singh has a total of 171 co-authors. Here are the top co-authors by number of shared publications:
- Christos Davatzikos
Artificial Intelligence in Biomedical Imaging Laboratory (AIBIL), Center for AI and Data Science for Integrated Diagnostics (AI2D), Perelman School of Medicine, University of Pennsylvania, Philadelphia, USA
9 shared publications - Aleks Stolicyn
Division of Psychiatry, Centre for Clinical Brain Sciences, University of Edinburgh, Edinburgh, UK
5 shared publications - Andrew Mark McIntosh
Royal Infirmary Edinburgh
5 shared publications - Cynthia H. Y. Fu
King’s College London
5 shared publications Dongtao Wei
School of Psychology, Southwest University, Chongqing, China
5 shared publicationsJiang Qiu
School of Psychology, Southwest University, Chongqing, China
5 shared publications
Latest Publications
- Open Access
- 13 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Personalized machine learning-guided radiation dose escalation in newly diagnosed glioblastoma: prospective pilot study
Glioblastoma is a highly aggressive primary brain tumor with near-universal recurrence despite maximal safe resection followed by standard chemoradiation. We conducted a prospective pilot study (ClinicalTrials.gov identifier: NCT03477513) with predefined endpoints and structured dose-escalation criteria to evaluate the feasibility and safety of personalized precision radiation therapy (PPRT) guided by machine learning (ML)-based maps of tumor infiltration. Twenty patients with newly diagnosed IDH-wildtype glioblastoma who underwent gross total resection received PPRT with concomitant and adjuvant temozolomide. The primary outcomes were median progression-free survival (PFS) and safety. Secondary outcomes included patterns of recurrence, rate of clinically significant toxicity, and median overall survival (OS). PPRT was feasible and well tolerated, with no grade ≥3 acute adverse events; 47% experienced grade 1 and 53% grade 2 events. Following therapy, radiation necrosis occurred in 47% of PPRT-temozolomide patients versus 12% in standard-of-care patients (p < 0.001). Median PFS was 24.4 months versus 11.6 months in 68 propensity score matched (PSM) historical control group (HR 0.28, 95% CI 0.13-0.61; p = 0.001). Median OS was 35.4 versus 17.7 months (HR 0.34, 95% CI 0.17-0.69; p = 0.003). Exploratory post hoc comparison suggests improved survival, but these findings are preliminary and require validation in randomized trials.
- 13 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessPersonalized machine learning-guided radiation dose escalation in newly diagnosed glioblastoma: prospective pilot study
Glioblastoma is a highly aggressive primary brain tumor with near-universal recurrence despite maximal safe resection followed by standard chemoradiation. We conducted a prospective pilot study (ClinicalTrials.gov identifier: NCT03477513) with predefined endpoints and structured dose-escalation criteria to evaluate the feasibility and safety of personalized precision radiation therapy (PPRT) guided by machine learning (ML)-based maps of tumor infiltration. Twenty patients with newly diagnosed IDH-wildtype glioblastoma who underwent gross total resection received PPRT with concomitant and adjuvant temozolomide. The primary outcomes were median progression-free survival (PFS) and safety. Secondary outcomes included patterns of recurrence, rate of clinically significant toxicity, and median overall survival (OS). PPRT was feasible and well tolerated, with no grade ≥3 acute adverse events; 47% experienced grade 1 and 53% grade 2 events. Following therapy, radiation necrosis occurred in 47% of PPRT-temozolomide patients versus 12% in standard-of-care patients (p < 0.001). Median PFS was 24.4 months versus 11.6 months in 68 propensity score matched (PSM) historical control group (HR 0.28, 95% CI 0.13-0.61; p = 0.001). Median OS was 35.4 versus 17.7 months (HR 0.34, 95% CI 0.17-0.69; p = 0.003). Exploratory post hoc comparison suggests improved survival, but these findings are preliminary and require validation in randomized trials.
- Open Access
- 15 November 2025
- Published by Springer Nature in Communications Medicine
- Vol. 5 (1)
Neuroanatomical dimensions in major depression linked to cognition, adverse life events, self-harm, metabolomics and genetics
Background Major depressive disorder (MDD) is a leading cause of disability worldwide, yet its diagnosis relies on clinical symptoms alone. Methods Using the semi-supervised machine learning algorithm, Heterogeneity through Discriminative Analysis (HYDRA), we had identified two neuroanatomical dimensions in deeply phenotyped (i.e., comprehensively assessed across neuroimaging, clinical, and behavioural domains), medication-free participants with MDD from the COORDINATE-MDD consortium. In the present study, we apply this pre-trained HYDRA model to the UK Biobank (UKB) to validate these dimensions in a large general population and a subsample with current depressive symptoms. Results Dimension 2 (D2), compared to Dimension 1 (D1), is characterized by reduced grey and white matter volumes and limited treatment response to antidepressant and placebo medications. Out-of-sample validation in the UKB general population (n = 37,235) confirms these neuroanatomical features and reveals D2 associations with cognitive impairments, adverse life events, self-harm and suicide attempts, a pro-atherogenic lipid profile, and genetic links to neurodegenerative traits. Similar profiles are observed in the UKB subsample with current depressive symptoms (n = 1455). Conclusions D1 and D2 represent distinct neurobiological mechanisms underlying MDD. The validation in a general population-based cohort and in a cohort sample with depressive symptoms delineates mechanisms underlying heterogeneity in MDD.
- 15 November 2025
- Published by Springer Nature in Communications Medicine
- Vol. 5 (1)
Open AccessNeuroanatomical dimensions in major depression linked to cognition, adverse life events, self-harm, metabolomics and genetics
Background Major depressive disorder (MDD) is a leading cause of disability worldwide, yet its diagnosis relies on clinical symptoms alone. Methods Using the semi-supervised machine learning algorithm, Heterogeneity through Discriminative Analysis (HYDRA), we had identified two neuroanatomical dimensions in deeply phenotyped (i.e., comprehensively assessed across neuroimaging, clinical, and behavioural domains), medication-free participants with MDD from the COORDINATE-MDD consortium. In the present study, we apply this pre-trained HYDRA model to the UK Biobank (UKB) to validate these dimensions in a large general population and a subsample with current depressive symptoms. Results Dimension 2 (D2), compared to Dimension 1 (D1), is characterized by reduced grey and white matter volumes and limited treatment response to antidepressant and placebo medications. Out-of-sample validation in the UKB general population (n = 37,235) confirms these neuroanatomical features and reveals D2 associations with cognitive impairments, adverse life events, self-harm and suicide attempts, a pro-atherogenic lipid profile, and genetic links to neurodegenerative traits. Similar profiles are observed in the UKB subsample with current depressive symptoms (n = 1455). Conclusions D1 and D2 represent distinct neurobiological mechanisms underlying MDD. The validation in a general population-based cohort and in a cohort sample with depressive symptoms delineates mechanisms underlying heterogeneity in MDD.
- 1 November 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- Vol. 7 (6) , e240299
DLMUSE: Robust Brain Segmentation in Seconds Using Deep Learning
The DLMUSE model was developed using a large, diverse dataset, achieving rapid brain MRI segmentation in seconds with performance comparable to state-of-the-art methods.
- 1 November 2025
- Published by Radiological Society of North America (RSNA) in Radiology: Artificial Intelligence
- Vol. 7 (6) , e240299
DLMUSE: Robust Brain Segmentation in Seconds Using Deep Learning
The DLMUSE model was developed using a large, diverse dataset, achieving rapid brain MRI segmentation in seconds with performance comparable to state-of-the-art methods.
- Open Access
- 26 March 2025
- Published by Springer Nature in Research Square
Neuroanatomical dimensions in major depression: external validation and links with cognition, adverse life events, self-harm, metabolomics and genetics
Major depressive disorder (MDD) is a leading cause of disability worldwide, yet its diagnosis relies on clinical symptoms alone. Using machine learning applied to deeply phenotyped, medication-free participants with MDD, we identified two neuroanatomical dimensions. Dimension 2 (D2), compared to Dimension 1 (D1), was characterized by reductions in grey and white matter and was associated with limited treatment response to both antidepressant and placebo medications. Validation in UK Biobank general population cohort (n = 37,235) confirmed that D2 is characterized by reduced grey and white matter, alongside widespread cognitive impairments, adverse events in both adulthood and childhood, increased self-harm and suicide attempts, a pro-atherogenic lipid profile, and genetic associations with neurodegenerative traits. These findings suggest that D1 and D2 reflect distinct neurobiological mechanisms underlying MDD, with important implications for and treatment outcomes. External validation was demonstrated in a general population-based cohort that delineated mechanisms underlying heterogeneity in MDD, identifying potential biomarkers that could aid in personalising treatment approaches for this debilitating disorder.
- 26 March 2025
- Published by Springer Nature in Research Square
Open AccessNeuroanatomical dimensions in major depression: external validation and links with cognition, adverse life events, self-harm, metabolomics and genetics
Major depressive disorder (MDD) is a leading cause of disability worldwide, yet its diagnosis relies on clinical symptoms alone. Using machine learning applied to deeply phenotyped, medication-free participants with MDD, we identified two neuroanatomical dimensions. Dimension 2 (D2), compared to Dimension 1 (D1), was characterized by reductions in grey and white matter and was associated with limited treatment response to both antidepressant and placebo medications. Validation in UK Biobank general population cohort (n = 37,235) confirmed that D2 is characterized by reduced grey and white matter, alongside widespread cognitive impairments, adverse events in both adulthood and childhood, increased self-harm and suicide attempts, a pro-atherogenic lipid profile, and genetic associations with neurodegenerative traits. These findings suggest that D1 and D2 reflect distinct neurobiological mechanisms underlying MDD, with important implications for and treatment outcomes. External validation was demonstrated in a general population-based cohort that delineated mechanisms underlying heterogeneity in MDD, identifying potential biomarkers that could aid in personalising treatment approaches for this debilitating disorder.
- Open Access
- 26 July 2024
- Published by Springer Nature in Nature Mental Health
- Vol. 2 (9) , 1120
Publisher Correction: Neuroanatomical dimensions in medication-free individuals with major depressive disorder and treatment response to SSRI antidepressant medications or placebo
- 26 July 2024
- Published by Springer Nature in Nature Mental Health
- Vol. 2 (9) , 1120
Open AccessPublisher Correction: Neuroanatomical dimensions in medication-free individuals with major depressive disorder and treatment response to SSRI antidepressant medications or placebo
- Open Access
- 12 January 2024
- Published by Springer Nature in Nature Mental Health
- Vol. 2 (2) , 164-176
Neuroanatomical dimensions in medication-free individuals with major depressive disorder and treatment response to SSRI antidepressant medications or placebo
Major depressive disorder (MDD) is a heterogeneous clinical syndrome with widespread subtle neuroanatomical correlates. Our objective was to identify the neuroanatomical dimensions that characterize MDD and predict treatment response to selective serotonin reuptake inhibitor (SSRI) antidepressants or placebo. In the COORDINATE-MDD consortium, raw MRI data were shared from international samples (N = 1,384) of medication-free individuals with first-episode and recurrent MDD (N = 685) in a current depressive episode of at least moderate severity, but not treatment-resistant depression, as well as healthy controls (N = 699). Prospective longitudinal data on treatment response were available for a subset of MDD individuals (N = 359). Treatments were either SSRI antidepressant medication (escitalopram, citalopram, sertraline) or placebo. Multi-center MRI data were harmonized, and HYDRA, a semi-supervised machine-learning clustering algorithm, was utilized to identify patterns in regional brain volumes that are associated with disease. MDD was optimally characterized by two neuroanatomical dimensions that exhibited distinct treatment responses to placebo and SSRI antidepressant medications. Dimension 1 was characterized by preserved gray and white matter (N = 290 MDD), whereas Dimension 2 was characterized by widespread subtle reductions in gray and white matter (N = 395 MDD) relative to healthy controls. Although there were no significant differences in age of onset, years of illness, number of episodes, or duration of current episode between dimensions, there was a significant interaction effect between dimensions and treatment response. Dimension 1 showed a significant improvement in depressive symptoms following treatment with SSRI medication (51.1%) but limited changes following placebo (28.6%). By contrast, Dimension 2 showed comparable improvements to either SSRI (46.9%) or placebo (42.2%) (β = –18.3, 95% CI (–34.3 to –2.3), P = 0.03). Findings from this case-control study indicate that neuroimaging-based markers can help identify the disease-based dimensions that constitute MDD and predict treatment response.
- 12 January 2024
- Published by Springer Nature in Nature Mental Health
- Vol. 2 (2) , 164-176
Open AccessNeuroanatomical dimensions in medication-free individuals with major depressive disorder and treatment response to SSRI antidepressant medications or placebo
Major depressive disorder (MDD) is a heterogeneous clinical syndrome with widespread subtle neuroanatomical correlates. Our objective was to identify the neuroanatomical dimensions that characterize MDD and predict treatment response to selective serotonin reuptake inhibitor (SSRI) antidepressants or placebo. In the COORDINATE-MDD consortium, raw MRI data were shared from international samples (N = 1,384) of medication-free individuals with first-episode and recurrent MDD (N = 685) in a current depressive episode of at least moderate severity, but not treatment-resistant depression, as well as healthy controls (N = 699). Prospective longitudinal data on treatment response were available for a subset of MDD individuals (N = 359). Treatments were either SSRI antidepressant medication (escitalopram, citalopram, sertraline) or placebo. Multi-center MRI data were harmonized, and HYDRA, a semi-supervised machine-learning clustering algorithm, was utilized to identify patterns in regional brain volumes that are associated with disease. MDD was optimally characterized by two neuroanatomical dimensions that exhibited distinct treatment responses to placebo and SSRI antidepressant medications. Dimension 1 was characterized by preserved gray and white matter (N = 290 MDD), whereas Dimension 2 was characterized by widespread subtle reductions in gray and white matter (N = 395 MDD) relative to healthy controls. Although there were no significant differences in age of onset, years of illness, number of episodes, or duration of current episode between dimensions, there was a significant interaction effect between dimensions and treatment response. Dimension 1 showed a significant improvement in depressive symptoms following treatment with SSRI medication (51.1%) but limited changes following placebo (28.6%). By contrast, Dimension 2 showed comparable improvements to either SSRI (46.9%) or placebo (42.2%) (β = –18.3, 95% CI (–34.3 to –2.3), P = 0.03). Findings from this case-control study indicate that neuroimaging-based markers can help identify the disease-based dimensions that constitute MDD and predict treatment response.
- Open Access
- 25 December 2023
- Published by Wiley in Alzheimer's & Dementia
- Vol. 19 (S17)
Artificial intelligence reveals brain aging patterns in 27,402 individuals without diagnosed cognitive impairment that are linked to genetics, biomedical measures, and cognitive decline
Background: Understanding heterogeneity of structural brain changes in aging may provide insights into susceptibility to neurodegenerative diseases. We characterize the genetics underlying brain structural heterogeneity within cognitively unimpaired (CU) individuals using data‐driven machine learning applied to a diverse dataset of 27,402 individuals from 11 neuroimaging studies from the iSTAGING consortium.Method: Structural brain morphologic patterns of CU individuals were independently examined in four decade‐long intervals spanning ages 45 to 85. Within each interval, Smile‐GAN (Yang et al., 2021) was trained on baseline anatomic and white matter hyperintensity (WMH) volumes. Smile‐GAN probability scores were used as phenotypes in genome‐wide association studies (GWAS). Specifically, we performed multiple linear regressions controlling for confounders (e.g., age) via Plink (Purcell et al., 2007). We observed longitudinal clustering stability across decades, so individuals from adjacent age groups were combined into broader age groups ([45,65), [65,85)) due to the large sample requirement of GWAS. Genomic loci, represented by the top leading single nucleotide polymorphisms (SNPs), were defined considering linkage disequilibrium. We investigated associations of SNPs with clinical traits and mapped them to genes using the GWAS Catalog (Buniello et al., 2019).Result: Three structural brain aging patterns, relative to resilient agers (A0), consistent across decades, emerged: A1, or ‘typical’ aging with low atrophy and WMHs, and two ‘advanced’ aging patterns, one showing elevated WMHs and modest atrophy (A2) and the other displaying severe, widespread atrophy and moderate WMH load (A3) (Figure 1). GWAS discovered eight and six genomic loci in [45,65) and [65,85) age groups, respectively (Table 1, Figure 2). The lead SNPs for A1 and A2 were previously associated with several cardiometabolic risk factors, WMHs, and regional brain volumes. Interestingly, rs4843552, previously associated with white matter microstructure and regional brain volumes, showed opposite effects for A1 (protection) and A2 (risk), consistent with the neuroimaging patterns. WDR41 in A3 group was previously associated with the age of Alzheimer’s disease onset (Herold et al., 2016).Conclusion: Reproducible neuroimaging patterns defined by regional atrophy and WMH burden were identified across CU individuals and demonstrated unique genetics. Further research is needed to elucidate the neuropathological pathways that mediate these relationships.
- 25 December 2023
- Published by Wiley in Alzheimer's & Dementia
- Vol. 19 (S17)
Open AccessArtificial intelligence reveals brain aging patterns in 27,402 individuals without diagnosed cognitive impairment that are linked to genetics, biomedical measures, and cognitive decline
Background: Understanding heterogeneity of structural brain changes in aging may provide insights into susceptibility to neurodegenerative diseases. We characterize the genetics underlying brain structural heterogeneity within cognitively unimpaired (CU) individuals using data‐driven machine learning applied to a diverse dataset of 27,402 individuals from 11 neuroimaging studies from the iSTAGING consortium.Method: Structural brain morphologic patterns of CU individuals were independently examined in four decade‐long intervals spanning ages 45 to 85. Within each interval, Smile‐GAN (Yang et al., 2021) was trained on baseline anatomic and white matter hyperintensity (WMH) volumes. Smile‐GAN probability scores were used as phenotypes in genome‐wide association studies (GWAS). Specifically, we performed multiple linear regressions controlling for confounders (e.g., age) via Plink (Purcell et al., 2007). We observed longitudinal clustering stability across decades, so individuals from adjacent age groups were combined into broader age groups ([45,65), [65,85)) due to the large sample requirement of GWAS. Genomic loci, represented by the top leading single nucleotide polymorphisms (SNPs), were defined considering linkage disequilibrium. We investigated associations of SNPs with clinical traits and mapped them to genes using the GWAS Catalog (Buniello et al., 2019).Result: Three structural brain aging patterns, relative to resilient agers (A0), consistent across decades, emerged: A1, or ‘typical’ aging with low atrophy and WMHs, and two ‘advanced’ aging patterns, one showing elevated WMHs and modest atrophy (A2) and the other displaying severe, widespread atrophy and moderate WMH load (A3) (Figure 1). GWAS discovered eight and six genomic loci in [45,65) and [65,85) age groups, respectively (Table 1, Figure 2). The lead SNPs for A1 and A2 were previously associated with several cardiometabolic risk factors, WMHs, and regional brain volumes. Interestingly, rs4843552, previously associated with white matter microstructure and regional brain volumes, showed opposite effects for A1 (protection) and A2 (risk), consistent with the neuroimaging patterns. WDR41 in A3 group was previously associated with the age of Alzheimer’s disease onset (Herold et al., 2016).Conclusion: Reproducible neuroimaging patterns defined by regional atrophy and WMH burden were identified across CU individuals and demonstrated unique genetics. Further research is needed to elucidate the neuropathological pathways that mediate these relationships.
- 16 September 2023
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 242, 107812
Imaging phenotypes predict overall survival in glioma more accurate than basic demographic and cell mutation profiles
- 16 September 2023
- Published by Elsevier in Computer Methods and Programs in Biomedicine
- Vol. 242, 107812
Imaging phenotypes predict overall survival in glioma more accurate than basic demographic and cell mutation profiles
- Open Access
- 22 December 2022
- Published by Springer Nature in Research Square
Neuroanatomical dimensions in medication-free individuals with major depressive disorder and treatment response to SSRI antidepressant medications or placebo
Importance: Major depressive disorder (MDD) is a heterogeneous clinical syndrome with widespread subtle neuroanatomical correlates. Identifying neuroimaging-based biomarkers might aid in defining the disease-related dimensions that characterize MDD and predict treatment response. Objective: To investigate the neuroanatomical dimensions that characterize MDD and predict treatment response to selective serotonin reuptake inhibitor (SSRI) antidepressant or placebo. Design: Big data consortium (COORDINATE-MDD) sharing raw MRI data in first episode and recurrent MDD, deep clinical phenotyping, and state-of-the art machine learning analysis, involving harmonization of multi-center MRI data and the application of semi-supervised machine learning clustering, HYDRA, to regional brain volumes. Setting: International, multi-center, community-based MDD and healthy controls. Participants: International sample (N=1384), consisting of medication-free, first episode and recurrent MDD individuals (N=685) in a current depressive episode of moderate to severe intensity, that is not treatment resistant depression, and healthy controls (N=699). Prospective longitudinal treatment response data were available in a subset of MDD individuals (N=359 MDD). Treatments were SSRI antidepressant medication (escitalopram, citalopram, sertraline) or placebo. Treatment duration was 6-8 weeks, and symptom severity was measured by clinician-rated scales. Main outcomes: First episode and recurrent MDD is optimally characterized by two neuroanatomical dimensions, which show distinct treatment effects to placebo and SSRI antidepressant medications. Results: Dimension 1 is characterized by preserved gray and white matter (N=290 MDD), whereas Dimension 2 is characterized by widespread subtle reductions in gray and white matter (N=395 MDD) relative to healthy controls. There are no significant differences in age of onset, years of illness, number of episodes, or duration of current episode between dimensions, but there is a significant dimension by treatment response interaction effect. Dimension 1 shows a significant decrease in depressive symptoms following treatment with SSRI medication (51.1%) but limited changes following placebo (28.6%), whereas Dimension 2 shows a comparable improvement to either SSRI (46.9%) or placebo (42.2%) (β=-18.3, 95% CI (-0.34 to -0.2), p=0.03). Conclusions and Relevance: Neuroimaging-based markers may aid in characterizing the MDD dimensions that predict treatment response. In an iterative process, we can characterize the disease-based dimensions that comprise MDD.
- 22 December 2022
- Published by Springer Nature in Research Square
Open AccessNeuroanatomical dimensions in medication-free individuals with major depressive disorder and treatment response to SSRI antidepressant medications or placebo
Importance: Major depressive disorder (MDD) is a heterogeneous clinical syndrome with widespread subtle neuroanatomical correlates. Identifying neuroimaging-based biomarkers might aid in defining the disease-related dimensions that characterize MDD and predict treatment response. Objective: To investigate the neuroanatomical dimensions that characterize MDD and predict treatment response to selective serotonin reuptake inhibitor (SSRI) antidepressant or placebo. Design: Big data consortium (COORDINATE-MDD) sharing raw MRI data in first episode and recurrent MDD, deep clinical phenotyping, and state-of-the art machine learning analysis, involving harmonization of multi-center MRI data and the application of semi-supervised machine learning clustering, HYDRA, to regional brain volumes. Setting: International, multi-center, community-based MDD and healthy controls. Participants: International sample (N=1384), consisting of medication-free, first episode and recurrent MDD individuals (N=685) in a current depressive episode of moderate to severe intensity, that is not treatment resistant depression, and healthy controls (N=699). Prospective longitudinal treatment response data were available in a subset of MDD individuals (N=359 MDD). Treatments were SSRI antidepressant medication (escitalopram, citalopram, sertraline) or placebo. Treatment duration was 6-8 weeks, and symptom severity was measured by clinician-rated scales. Main outcomes: First episode and recurrent MDD is optimally characterized by two neuroanatomical dimensions, which show distinct treatment effects to placebo and SSRI antidepressant medications. Results: Dimension 1 is characterized by preserved gray and white matter (N=290 MDD), whereas Dimension 2 is characterized by widespread subtle reductions in gray and white matter (N=395 MDD) relative to healthy controls. There are no significant differences in age of onset, years of illness, number of episodes, or duration of current episode between dimensions, but there is a significant dimension by treatment response interaction effect. Dimension 1 shows a significant decrease in depressive symptoms following treatment with SSRI medication (51.1%) but limited changes following placebo (28.6%), whereas Dimension 2 shows a comparable improvement to either SSRI (46.9%) or placebo (42.2%) (β=-18.3, 95% CI (-0.34 to -0.2), p=0.03). Conclusions and Relevance: Neuroimaging-based markers may aid in characterizing the MDD dimensions that predict treatment response. In an iterative process, we can characterize the disease-based dimensions that comprise MDD.
- Open Access
- 15 August 2021
- Published by MDPI AG in Applied Sciences
- Vol. 11 (16) , 7488
Interactive Machine Learning-Based Multi-Label Segmentation of Solid Tumors and Organs
We seek the development and evaluation of a fast, accurate, and consistent method for general-purpose segmentation, based on interactive machine learning (IML). To validate our method, we identified retrospective cohorts of 20 brain, 50 breast, and 50 lung cancer patients, as well as 20 spleen scans, with corresponding ground truth annotations. Utilizing very brief user training annotations and the adaptive geodesic distance transform, an ensemble of SVMs is trained, providing a patient-specific model applied to the whole image. Two experts segmented each cohort twice with our method and twice manually. The IML method was faster than manual annotation by 53.1% on average. We found significant (p < 0.001) overlap difference for spleen (DiceIML/DiceManual = 0.91/0.87), breast tumors (DiceIML/DiceManual = 0.84/0.82), and lung nodules (DiceIML/DiceManual = 0.78/0.83). For intra-rater consistency, a significant (p = 0.003) difference was found for spleen (DiceIML/DiceManual = 0.91/0.89). For inter-rater consistency, significant (p < 0.045) differences were found for spleen (DiceIML/DiceManual = 0.91/0.87), breast (DiceIML/DiceManual = 0.86/0.81), lung (DiceIML/DiceManual = 0.85/0.89), the non-enhancing (DiceIML/DiceManual = 0.79/0.67) and the enhancing (DiceIML/DiceManual = 0.79/0.84) brain tumor sub-regions, which, in aggregation, favored our method. Quantitative evaluation for speed, spatial overlap, and consistency, reveals the benefits of our proposed method when compared with manual annotation, for several clinically relevant problems. We publicly release our implementation through CaPTk (Cancer Imaging Phenomics Toolkit) and as an MITK plugin.
- 15 August 2021
- Published by MDPI AG in Applied Sciences
- Vol. 11 (16) , 7488
Open AccessInteractive Machine Learning-Based Multi-Label Segmentation of Solid Tumors and Organs
We seek the development and evaluation of a fast, accurate, and consistent method for general-purpose segmentation, based on interactive machine learning (IML). To validate our method, we identified retrospective cohorts of 20 brain, 50 breast, and 50 lung cancer patients, as well as 20 spleen scans, with corresponding ground truth annotations. Utilizing very brief user training annotations and the adaptive geodesic distance transform, an ensemble of SVMs is trained, providing a patient-specific model applied to the whole image. Two experts segmented each cohort twice with our method and twice manually. The IML method was faster than manual annotation by 53.1% on average. We found significant (p < 0.001) overlap difference for spleen (DiceIML/DiceManual = 0.91/0.87), breast tumors (DiceIML/DiceManual = 0.84/0.82), and lung nodules (DiceIML/DiceManual = 0.78/0.83). For intra-rater consistency, a significant (p = 0.003) difference was found for spleen (DiceIML/DiceManual = 0.91/0.89). For inter-rater consistency, significant (p < 0.045) differences were found for spleen (DiceIML/DiceManual = 0.91/0.87), breast (DiceIML/DiceManual = 0.86/0.81), lung (DiceIML/DiceManual = 0.85/0.89), the non-enhancing (DiceIML/DiceManual = 0.79/0.67) and the enhancing (DiceIML/DiceManual = 0.79/0.84) brain tumor sub-regions, which, in aggregation, favored our method. Quantitative evaluation for speed, spatial overlap, and consistency, reveals the benefits of our proposed method when compared with manual annotation, for several clinically relevant problems. We publicly release our implementation through CaPTk (Cancer Imaging Phenomics Toolkit) and as an MITK plugin.