Spectral decoupling for training transferable neural networks in medical imaging

Pohjonen, Joona; Stürenberg, Carolin; Rannikko, Antti; Mirtti, Tuomas; Pitkänen, Esa

doi:10.1016/j.isci.2022.103767

Cited by 7 publications

(5 citation statements)

References 22 publications

(34 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, when we tested the methods on somatic variant calls before quality filtering, we discovered that all the tested machine learning models, including MuAt, were sensitive to variant calling artefacts. This finding builds upon previous reports of fragility in ML models and emphasizes the importance of using highquality data and creating ML models that are robust to input data variability resulting from differences in data generating processes [57].…”

Section: Discussionsupporting

confidence: 81%

“…Deep learning models have been reported to be fragile, or sensitive to small changes in input data, leading to incorrect or misleading conclusions drawn from model outputs [56,57]. As model fragility can be a significant challenge when deploying machine learning models for clinical use [58], we investigated whether MuAt models would be able to maintain robustness when faced with shifts in input data distribution.…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

Mutation-Attention (MuAt): deep representation learning of somatic mutations for tumour typing and subtyping

Sanjaya,

Maljanen,

Katainen

et al. 2023

Genome Med

Self Cite

View full text Add to dashboard Cite

Background Cancer genome sequencing enables accurate classification of tumours and tumour subtypes. However, prediction performance is still limited using exome-only sequencing and for tumour types with low somatic mutation burden such as many paediatric tumours. Moreover, the ability to leverage deep representation learning in discovery of tumour entities remains unknown. Methods We introduce here Mutation-Attention (MuAt), a deep neural network to learn representations of simple and complex somatic alterations for prediction of tumour types and subtypes. In contrast to many previous methods, MuAt utilizes the attention mechanism on individual mutations instead of aggregated mutation counts. Results We trained MuAt models on 2587 whole cancer genomes (24 tumour types) from the Pan-Cancer Analysis of Whole Genomes (PCAWG) and 7352 cancer exomes (20 types) from the Cancer Genome Atlas (TCGA). MuAt achieved prediction accuracy of 89% for whole genomes and 64% for whole exomes, and a top-5 accuracy of 97% and 90%, respectively. MuAt models were found to be well-calibrated and perform well in three independent whole cancer genome cohorts with 10,361 tumours in total. We show MuAt to be able to learn clinically and biologically relevant tumour entities including acral melanoma, SHH-activated medulloblastoma, SPOP-associated prostate cancer, microsatellite instability, POLE proofreading deficiency, and MUTYH-associated pancreatic endocrine tumours without these tumour subtypes and subgroups being provided as training labels. Finally, scrunity of MuAt attention matrices revealed both ubiquitous and tumour-type specific patterns of simple and complex somatic mutations. Conclusions Integrated representations of somatic alterations learnt by MuAt were able to accurately identify histological tumour types and identify tumour entities, with potential to impact precision cancer medicine.

show abstract

Section: Discussionsupporting

confidence: 81%

Section: Discussionmentioning

confidence: 99%

Mutation-Attention (MuAt): deep representation learning of somatic mutations for tumour typing and subtyping

Sanjaya,

Maljanen,

Katainen

et al. 2023

Genome Med

Self Cite

View full text Add to dashboard Cite

show abstract

“… 28 detection and automated GG grading CNN, semi-supervised learning biopsy training: Local (580 slides), PANDA (Radboud) dataset external: PANDA (Karolinska) dataset detection: AUC: 0.92 GG grading # : accuracy: 0.831; κ quad 0.93 GG2 vs. GG3-5: AUC: 0.93 Pohjonen et al. 29 improve generalization problem for the detection of prostate cancer neural network trained with spectral decoupling – training: 90 patients from Helsinki external: PESO dataset networks trained with spectral decoupling achieve up to 9.5% point higher accuracy on external datasets (The author did not report the exact value of accuracy.) PANDA challenge 30 detection and automated GG grading various DL algorithms biopsy training: PANDA dataset external: two cohorts (714 and 330 slides) detection: sensitivity 0.986 and 0.0.977; specificity 0.752 and 0.843 GG grading # : κ quad 0.862 and 0.868 Silva-Rodríguez et al.…”

Section: Development Of Ai Models For Prostate Cancer Managementmentioning

confidence: 99%

“… 30 , 50 Heterogeneity can stem from various factors, such as staining variations, artifacts, and imaging differences between scanners. 29 , 75 , 106 , 107 To overcome this, an ideal approach is to have a sufficiently large and diverse training set, such as continuously collecting all cases over a certain period of time from multiple institutions, in order to cover all possible variations in the real world to represent the entire target population. From the perspective of fully utilizing the existing data, data augmentation techniques such as rotation, flipping, and color enhancement can be applied to enhance the original training set.…”

Section: Challenges Of Application Of Ai In Clinicmentioning

confidence: 99%

Harnessing artificial intelligence for prostate cancer management

Zhu,

Pan,

Mou

et al. 2024

Cell Reports Medicine

View full text Add to dashboard Cite

“…Spectral decoupling, weight normalization, and L1 penalty have been applied to varying degrees in studies concerning computational pathologyonce [1], twice [2,3], and several times [4], respectively. As reported in general use cases, these methods for inducing sparsity tend to improve overall performance.…”

Section: Introductionmentioning

confidence: 99%

The effects of sparsity induction methods on attention-based multiple instance learning applied to Camelyon16

Tavolara

Gürcan²,

Niazi

2023

Medical Imaging 2023: Digital and Computational Pathology

View full text Add to dashboard Cite

Spectral decoupling, weight normalization, and L1 loss have been applied to varying degrees in studies concerning computational pathology. These methods induce sparsity and tend to improve overall model performance. However, no work has combined these methods to improve model performance. The work presented here combines and compares these three methods in an attention-based multiple instance learning model to classify whole slide histopathology images from Camelyon16. We observed that spectral decoupling improves accuracy and area under the curve (AUC), but that weight normalization and L1 loss do not. However, when either of the latter is combined with spectral decoupling, accuracy, and AUC further improve over just spectral decoupling. Finally, we demonstrate that varying the magnitude with which these three methods affect model training considerably affects the resulting testing accuracy and AUC.

show abstract

Spectral decoupling for training transferable neural networks in medical imaging

Cited by 7 publications

References 22 publications

Mutation-Attention (MuAt): deep representation learning of somatic mutations for tumour typing and subtyping

Mutation-Attention (MuAt): deep representation learning of somatic mutations for tumour typing and subtyping

Harnessing artificial intelligence for prostate cancer management

The effects of sparsity induction methods on attention-based multiple instance learning applied to Camelyon16

Contact Info

Product

Resources

About