Generalizable Feature Learning in the Presence of Data Bias and Domain Class Imbalance with Application to Skin Lesion Classification

Yoon, Chris J. M.; Hamarneh, Ghassan; Garbi, Rafeef

doi:10.1007/978-3-030-32251-9_40

Cited by 35 publications

(24 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Second, the training, validation, and testing partitions for Brinker et al all come from the same data source, i.e., the ISIC Archive, whereas our model was trained on the Interactive Atlas of Dermoscopy and evaluated on images from the ISIC Archive, leading to a domain shift. CNNs have been shown to exhibit poor generalizability for skin lesion classification tasks when trained and evaluated on separate datasets 34 . Despite this, our multi-task prediction model is able to adapt to the new domain and exhibits strong generalization performance for clinical management predictions.…”

Section: Resultsmentioning

confidence: 99%

Predicting the Clinical Management of Skin Lesions using Deep Learning

Abhishek

Kawahara

Hamarneh

2020

Preprint

Self Cite

View full text Add to dashboard Cite

Automated machine learning approaches to skin lesion diagnosis from images are approaching dermatologist-level performance. However, current machine learning approaches that suggest management decisions rely on predicting the underlying skin condition to infer a management decision without considering the variability of management decisions that may exist within a single condition. We present the first work to explore image-based prediction of clinical management decisions directly without explicitly predicting the diagnosis. In particular, we use clinical and dermoscopic images of skin lesions along with patient metadata from the Interactive Atlas of Dermoscopy dataset (1,011 cases; 20 disease labels; 3 management decisions) and demonstrate that predicting management labels directly is more accurate than predicting the diagnosis and then inferring the management decision (11.74% and 9.07% improvement in overall accuracy and AUROC respectively), statistically significant at p < 0.001. Directly predicting management decisions also considerably reduces the over-excision rate as compared to management decisions inferred from diagnosis predictions (24.56% fewer cases wrongly predicted to be excised). Furthermore, we show that training a model to also simultaneously predict the seven-point criteria and the diagnosis of skin lesions yields an even higher accuracy (improvements of 5.44% and 0.73% in overall accuracy and AUROC respectively) of management predictions. Finally, we demonstrate our model's generalizability by evaluating on the publicly available MClass-D dataset and show that our model agrees with the clinical management recommendations of 157 dermatologists as much as they agree amongst each other.

show abstract

Section: Resultsmentioning

confidence: 99%

Predicting the Clinical Management of Skin Lesions using Deep Learning

Abhishek

Kawahara

Hamarneh

2020

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…Skin lesion dataset is a combination of 7 public datasets for skin lesion detection collected from different equipments. The main dataset is HAM10000 [50] which is used as part of the source data of all experiments following the setup of [51], [52]. The other datasets are Dermofit (DMF) [53], Derm7pt (D7P) [54], MSK [55], PH2 [56], SONIC (SON) [55], and UDA [55].…”

Section: A Datasetsmentioning

confidence: 99%

“…All the datasets contain 7 common lesions which are melanoma (mel), melanocytic nevus (nv), dermatofibroma (df), basal cell carcinoma (bcc), vascular lesion (vasc), benign keratosis (bkl), and actinic keratosis (akiec). Following [52] we split the data into training (50%), validation (20%) and testing set (30%) in a stratified manner. In each experiment we choose one of the secondary datasets as a target domain and keep HAM10000 and the other dataset for the source domains.…”

Section: A Datasetsmentioning

confidence: 99%

Adversarial Reconstruction Loss for Domain Generalization

et al. 2021

View full text Add to dashboard Cite

The biggest fear when deploying machine learning models to the real world is their ability to handle the new data. This problem is significant especially in medicine, where models trained on rich high-quality data extracted from large hospitals do not scale to small regional hospitals. One of the clinical challenges addressed in this work is magnetic resonance image generalization for improved visualization and diagnosis of hip abnormalities such as femoroacetabular impingement and dysplasia. Domain Generalization (DG) is a field in machine learning that tries to solve the model's dependency on the training data by leveraging many related but different data sources. We present a new method for DG that is both efficient and fast, unlike the most current state of art methods, which add a substantial computational burden making it hard to fine-tune. Our model trains an autoencoder setting on top of the classifier, but the encoder is trained on the adversarial reconstruction loss forcing it to forget style information while extracting features useful for classification. Our approach aims to force the encoder to generate domain-invariant representations that are still category informative by pushing it in both directions. Our method has proven universal and was validated on four different benchmarks for domain generalization, outperforming state of the art on RMNIST, VLCS and IXMAS with a 0.70% increase in accuracy and providing comparable results on PACS with a 0.02% difference. Our method was also evaluated for unsupervised domain adaptation and has shown to be quite an effective method against over-fitting.

show abstract

“…Domain alignment methods aim to learn domaininvariant representations of the data by aligning features across the source domains. The reduction of the representation distribution mismatch across source domains can be achieved by minimizing the maximum mean discrepancy criteria (Gretton et al 2012) combined with an adversarial autoencoder (MMD) (Li et al 2018b), minimizing the difference between the means (Tzeng et al 2014) or covariance matrices (CORAL) (Sun and Saenko 2016) in the embedding space across different domains, or minimizing a con-trastive loss as a regularization (Motiian et al 2017;Yoon, Hamarneh, and Garbi 2019;Mahajan, Tople, and Sharma 2020), e.g., SelfReg (Kim et al 2021). The domain alignment is also addressed by promoting the loss gradient alignment across different domains via inner product maximization (Fish) (Shi et al 2021) or binary (AND-mask) (Parascandolo et al 2020;Shahtalebi et al 2021) or continuous gradient masking (SAND-mask) (Shahtalebi et al 2021).…”

Section: Domain Generalizationmentioning

confidence: 99%

COLUMBUS: Automated Discovery of New Multi-Level Features for Domain Generalization via Knowledge Corruption

Frikha¹,

Krompaß²,

Tresp³

2021

Preprint

View full text Add to dashboard Cite

Machine learning models that can generalize to unseen domains are essential when applied in real-world scenarios involving strong domain shifts. We address the challenging domain generalization (DG) problem, where a model trained on a set of source domains is expected to generalize well in unseen domains without any exposure to their data. The main challenge of DG is that the features learned from the source domains are not necessarily present in the unseen target domains, leading to performance deterioration. We assume that learning a richer set of features is crucial to improve the transfer to a wider set of unknown domains. For this reason, we propose COLUMBUS, a method that enforces new feature discovery via a targeted corruption of the most relevant input and multi-level representations of the data. We conduct an extensive empirical evaluation to demonstrate the effectiveness of the proposed approach which achieves new state-of-the-art results by outperforming 18 DG algorithms on multiple DG benchmark datasets in the DOMAINBED framework.

show abstract

Generalizable Feature Learning in the Presence of Data Bias and Domain Class Imbalance with Application to Skin Lesion Classification

Cited by 35 publications

References 13 publications

Predicting the Clinical Management of Skin Lesions using Deep Learning

Predicting the Clinical Management of Skin Lesions using Deep Learning

Adversarial Reconstruction Loss for Domain Generalization

COLUMBUS: Automated Discovery of New Multi-Level Features for Domain Generalization via Knowledge Corruption

Contact Info

Product

Resources

About