Melanoma skin cancer is considered as one of the most common diseases in the world. Detecting such diseases at early stage is important to saving lives. During medical examinations, it is not an easy task to visually inspect such lesions, as there are similarities between lesions. Technological advances in the form of deep learning methods have been used for diagnosing skin lesions. Over the last decade, deep learning, especially CNN (convolutional neural networks), has been found one of the promising methods to achieve state-of-art results in a variety of medical imaging applications. However, ConvNets’ capabilities are considered limited due to the lack of understanding of long-range spatial relations in images. The recently proposed Vision Transformer (ViT) for image classification employs a purely self-attention-based model that learns long-range spatial relations to focus on the image’s relevant parts. To achieve better performance, existing transformer-based network architectures require large-scale datasets. However, because medical imaging datasets are small, applying pure transformers to medical image analysis is difficult. ViT emphasizes the low-resolution features, claiming that the successive downsampling results in a lack of detailed localization information, rendering it unsuitable for skin lesion image classification. To improve the recovery of detailed localization information, several ViT-based image segmentation methods have recently been combined with ConvNets in the natural image domain. This study provides a comprehensive comparative study of U-Net and attention-based methods for skin lesion image segmentation, which will assist in the diagnosis of skin lesions. The results show that the hybrid TransUNet, with an accuracy of 92.11% and dice coefficient of 89.84%, outperforms other benchmarking methods.

Modality corresponding to medical images is a vital filter in medical image retrieval systems. This article presents the classification of modalities of medical images based on the usage of principles of hyper-dimensional computing and reservoir computing. It is demonstrated that the highest classification accuracy of the proposed method is on a par with the best classical method for the given dataset (83% vs. 84%). The major positive property of the proposed method is that it does not require any optimization routine during the training phase and naturally allows for incremental learning upon the availability of new training data.

Modeling low level features to high level semantics in medical imaging is an important aspect in filtering anatomy objects. Bag of Visual Words (BOVW) representations have been proven effective to model these low level features to mid level representations. Convolutional neural nets are learning systems that can automatically extract high-quality representations from raw images. However, their deployment in the medical field is still a bit challenging due to the lack of training data. In this paper, learned features that are obtained by training convolutional neural networks are compared with our proposed hand-crafted HSIFT features. The HSIFT feature is a symmetric fusion of a Harris corner detector and the Scale Invariance Transform process (SIFT) with BOVW representation. The SIFT process is enhanced as well as the classification technique by adopting bagging with a surrogate split method. Quantitative evaluation shows that our proposed hand-crafted HSIFT feature outperforms the learned features from convolutional neural networks in discriminating anatomy image classes.

With the emergence of single-cell foundation models, an important question arises: how do these models perform when trained on datasets having an imbalance in cell type distribution due to rare cell types or biased sampling? We benchmark three foundation models, scGPT, scBERT, and Geneformer, using skewed single-cell cell-type distribution for cell-type annotation. While all models had reduced performance when challenged with rare cell types, scGPT and scBERT, performed better than Geneformer. Notably, in contrast to scGPT and scBERT, Geneformer uses ordinal positions of the tokenized genes rather than actual raw gene expression values. To mitigate the effect of a skewed distribution, we find that random oversampling, but not random undersampling, improved the performance for all three foundation models. Finally, scGPT, using FlashAttention, has the fastest computational speed, whereas scBERT is more memory-efficient. We conclude that tokenization and data representation are essential areas of research, and new strategies are needed to mitigate the effects of imbalanced learning in single-cell foundation models. Code and data for reproducibility are available at https://github.com/SabbaghCodes/ImbalancedLearningForSingleCellFoundationModels .