Towards interpretable machine learning models for diagnosis aid: A case study on attention deficit/hyperactivity disorder.

Attention Deficit/Hyperactivity Disorder (ADHD) is a neurodevelopmental disorder that has heavy consequences on a child's wellbeing, especially in the academic, psychological and relational planes. The current evaluation of the disorder is supported by clinical assessment and written tests. A definitive diagnosis is usually made based on the DSM-V criteria. There is a lot of ongoing research on ADHD, in order to determine the neurophysiological basis of the disorder and to reach a more objective diagnosis. The advent of Machine Learning (ML) opens up promising prospects for the development of systems able to predict a diagnosis from phenotypic and neuroimaging data. This was the reason why the ADHD-200 contest was launched a few years ago. Based on the publicly available ADHD-200 collection, participants were challenged to predict ADHD with the best possible predictive accuracy. In the present work, we propose instead a ML methodology which primarily places importance on the explanatory power of a model. Such an approach is intended to achieve a fair trade-off between the needs of performance and interpretability expected from medical diagnosis aid systems. We applied our methodology on a data sample extracted from the ADHD-200 collection, through the development of decision trees which are valued for their readability. Our analysis indicates the relevance of the limbic system for the diagnosis of the disorder. Moreover, while providing explanations that make sense, the resulting decision tree performs favorably given the recent results reported in the literature.

Three-dimensional spine model reconstruction using one-class SVM regularization.

Statistical shape models have become essential for medical image registration or segmentation and are used in many biomedical applications. These models are often based on Gaussian distributions learned from a training set. We propose in this paper a shape model which does not rely on the estimation of a Gaussian distribution, but on similarities computed with a kernel function. Our model takes advantage of the one-class support vector machine (OCSVM) to do so. In this context, we propose in this paper a method for reconstructing the spine of scoliotic patients using OCSVM regularization. Current state-of-the-art methods use conventional statistical shape models, and the reconstruction is commonly processed by minimizing a Mahalanobis distance. Nevertheless, when a shape differs significantly from the statistical model, the associated Mahalanobis distance often overstates the need for statistical regularization. We show that OCSVM regularization is more robust and is less sensitive to weak landmarks definition and is hardly influenced by the presence of outliers in the training data. The proposed OCSVM model applied to 3-D spine reconstruction was evaluated on real patient data, and results showed that our approach allows precise reconstruction.

Cervical spine mobility analysis on radiographs: a fully automatic approach.

Conventional X-ray radiography remains nowadays the most common method to analyze spinal mobility in two dimensions. Therefore, the objective of this paper is to develop a framework dedicated to the fully automatic cervical spine mobility analysis on X-ray images. To this aim, we propose an approach based on three main steps: fully automatic vertebra detection, vertebra segmentation and angular measurement. The accuracy of the method was assessed for a total of 245 vertebræ. For the vertebra detection, we proposed an adapted version of two descriptors, namely Scale-invariant Feature Transform (SIFT) and Speeded-up Robust Features (SURF), coupled with a multi-class Support Vector Machine (SVM) classifier. Vertebræ are successfully detected in 89.8% of cases and it is demonstrated that SURF slightly outperforms SIFT. The Active Shape Model approach was considered as a segmentation procedure. We observed that a statistical shape model specific to the vertebral level improves the results. Angular errors of cervical spine mobility are presented. We showed that these errors remain within the inter-operator variability of the reference method.

Fast 3D spine reconstruction of postoperative patients using a multilevel statistical model.

Severe cases of spinal deformities such as scoliosis are usually treated by a surgery where instrumentation (hooks, screws and rods) is installed to the spine to correct deformities. Even if the purpose is to obtain a normal spine curve, the result is often straighter than normal. In this paper, we propose a fast statistical reconstruction algorithm based on a general model which can deal with such instrumented spines. To this end, we present the concept of multilevel statistical model where the data are decomposed into a within-group and a between-group component. The reconstruction procedure is formulated as a second-order cone program which can be solved very fast (few tenths of a second). Reconstruction errors were evaluated on real patient data and results showed that multilevel modeling allows better 3D reconstruction than classical models.

A framework of vertebra segmentation using the active shape model-based approach.

We propose a medical image segmentation approach based on the Active Shape Model theory. We apply this method for cervical vertebra detection. The main advantage of this approach is the application of a statistical model created after a training stage. Thus, the knowledge and interaction of the domain expert intervene in this approach. Our application allows the use of two different models, that is, a global one (with several vertebrae) and a local one (with a single vertebra). Two modes of segmentation are also proposed: manual and semiautomatic. For the manual mode, only two points are selected by the user on a given image. The first point needs to be close to the lower anterior corner of the last vertebra and the second near the upper anterior corner of the first vertebra. These two points are required to initialize the segmentation process. We propose to use the Harris corner detector combined with three successive filters to carry out the semiautomatic process. The results obtained on a large set of X-ray images are very promising.

Heterogeneous computing for vertebra detection and segmentation in x-ray images.

The context of this work is related to the vertebra segmentation. The method we propose is based on the active shape model (ASM). An original approach taking advantage of the edge polygonal approximation was developed to locate the vertebra positions in a X-ray image. Despite the fact that segmentation results show good efficiency, the time is a key variable that has always to be optimized in a medical context. Therefore, we present how vertebra extraction can efficiently be performed in exploiting the full computing power of parallel (GPU) and heterogeneous (multi-CPU/multi-GPU) architectures. We propose a parallel hybrid implementation of the most intensive steps enabling to boost performance. Experimentations have been conducted using a set of high-resolution X-ray medical images, showing a global speedup ranging from 3 to 22, by comparison with the CPU implementation. Data transfer times between CPU and GPU memories were included in the execution times of our proposed implementation.