Machine Learning Techniques for Biomedical Image Segmentation: An Overview of Technical Aspects and Introduction to State-of-Art Applications (1911.02521v1)

Abstract: In recent years, significant progress has been made in developing more accurate and efficient machine learning algorithms for segmentation of medical and natural images. In this review article, we highlight the imperative role of machine learning algorithms in enabling efficient and accurate segmentation in the field of medical imaging. We specifically focus on several key studies pertaining to the application of machine learning methods to biomedical image segmentation. We review classical machine learning algorithms such as Markov random fields, k-means clustering, random forest, etc. Although such classical learning models are often less accurate compared to the deep learning techniques, they are often more sample efficient and have a less complex structure. We also review different deep learning architectures, such as the artificial neural networks (ANNs), the convolutional neural networks (CNNs), and the recurrent neural networks (RNNs), and present the segmentation results attained by those learning models that were published in the past three years. We highlight the successes and limitations of each machine learning paradigm. In addition, we discuss several challenges related to the training of different machine learning models, and we present some heuristics to address those challenges.

• The paper shows that kernel SVMs can outperform deep learning models in small sample settings, achieving higher mean IoU scores.
• It details the use of CNNs, FCNs, and U-Nets for robust pixel-wise classification and hierarchical feature extraction.
• The review emphasizes data augmentation and transfer learning as key strategies to overcome data scarcity in medical imaging.

Machine Learning Techniques for Biomedical Image Segmentation

The paper, "Machine Learning Techniques for Biomedical Image Segmentation: An Overview of Technical Aspects and Introduction to State-of-Art Applications," provides a comprehensive review of the application of machine learning algorithms to biomedical image segmentation. The authors systematically discuss both classical machine learning methods and contemporary deep learning approaches, highlighting their utility in medical image segmentation tasks.

Initially, the paper focuses on classical machine learning techniques such as Markov Random Fields (MRF), Support Vector Machines (SVMs), and random forests. These methods, while maintaining simplicity and lower computational burden, are evaluated for their capacity to handle segmentation tasks when trained with limited datasets. For instance, the kernel SVM method, which is complemented by a filter bank for feature extraction, shows promising adaptability in small sample environments, as evidenced by its superior performance in segmenting GIANA challenge images compared to Fully Convolutional Networks (FCNs).

The review section on deep learning expands on its transformational impact on image segmentation methodologies, explaining architectures like CNNs and RNNs, with particular attention to their implementation in medical segmentation tasks. A noteworthy emphasis is placed on CNNs' hierarchical feature learning, which allows robust translation and deformation invariance through layers of convolutional and pooling operations. The risk of overfitting these extensive models, particularly when trained on limited medical image datasets, is acknowledged, alongside strategies such as transfer learning and data augmentation to mitigate these risks.

The paper also covers FCNs and U-Nets, which are highlighted for their powerful segmentation capabilities through end-to-end training for pixel-wise classification tasks. U-Nets, distinguished by their skip connections between contracting and expanding paths, exemplify sophisticated use of hierarchical structures to enhance segmentation accuracy. Additionally, issues of data scarcity in biomedical datasets are addressed through techniques such as Generative Adversarial Networks (GANs) for synthetic dataset augmentation.

The results section underscores the effectiveness of kernel SVMs over deep learning networks like FCNs in settings constrained by data availability, demonstrating superior test outcomes in the form of mean Intersection over Union (IoU) scores. The observations made through t-SNE plots on kernel feature mapping elucidate critical aspects of feature representation capabilities, and the necessity of meticulous bandwidth parameter tuning to achieve optimal classification accuracy.

In terms of theoretical implications, the survey illustrates the paradigmatic shift towards deep learning in medical domains, where large volumes of data enable complex model training. However, it also brings to light the ongoing need for data-efficient models and refining existing algorithms for enhanced performance in resource-limited settings. This review provides a thorough insight into past and current methodologies, serving as a pivot for the development of more robust, scalable algorithms in the domain of medical image analysis.

Looking forward, the paper suggests exploring advancements in unsupervised and semi-supervised learning models, such as Graph Neural Networks (GNNs), which might provide enhanced capabilities for medical image segmentation tasks. Moreover, the framework of adaptive network structures, such as those explored in AdaNet, points towards a future where networks self-optimizing on-the-fly could be achieved, further bolstering the toolset for researchers in this field.