Brain malformation assessment is usually performed by visual inspection of MRI images by experienced neuroradiologists. Such approach is therefore subjective and provides only a qualitative description of the detected malformations. The purpose of this study is to exploit machine learning methods to automatically detect and quantitatively characterize corpus callosum (CC) malformations, improving the inter-subject diagnostic reproducibility.

Dataset: 128 subjects (Age mean and SD: 7.3 ± 1.9 yo; males/females: 85/43), including healthy, mentally retarded and syndromic subjects, were recruited. The MRI protocol included a T1-weighted sequence at a resolution of 1x1x1 mm³ in a 3T Philips Achieva scanner with a 32-channel head coil.

Image analysis: image preprocessing and segmentation were performed using FSL tools¹. The CC was extracted as in Herron et al² and geometric features were computed using home-made scripts. The feature set included the area, the perimeter length, the skeleton length and curvature, the splenium-rostrum distance, and the thickness profile. A one-class multiple-kernel Support Vector Machine (SVM) classifier was used to assess the CC diagnosis. The classifier assigned a label to each object (i.e. “normal-shaped”/”malformed” CC), but did not provide any information about what led to the classification. This problem was overcome by providing a new regularized Discriminative Direction (DD) analysis³, which was applied to multiple-kernel SMV classifiers⁴. The DD analysis computed the minimum transformations required to change a “malformed” CC into a “normal-shaped” one. Consequently, the DD analysis provided a quantitative assessment of the malformation pattern and severity.

Inter-subject reproducibility experiment: four neuroradiologists, with more than 10 years of experience in pediatric neuroimaging, were asked to classify on a visual base each CC as “normal-shaped” or “malformed”, being blind about the patient clinical picture (“blind-diagnoses”). The “normal-shaped” CCs from the consensus among three neuroradiologists were used to train the SVM classifier. One year later, all physicians were asked to repeat the classification of the same dataset using the classifier output as Assisted Diagnosis Tool (ADT) (“ADT-diagnoses”). Agreement rates were computed for each couple of physicians to assess the inter-subject reproducibility.

Figure 1 reports an example of the ADT output. A large panel shows the middle sagittal slice of the T1 image where the CC is located, while several smaller panels in the lower part of the image report the classifier diagnosis and DD analysis. The latter is reported both with a graphical representation of some features (e.g. thickness profile or skeleton curvature) and with a numeric assessment of the malformation degree for the others (e.g. area, perimeter length, etc). A color and intensity code was used to visualize the versus and the magnitude of the malformation, respectively. Red color is used to indicate the features that are too small with respect to a normal-shaped CC. Conversely, blue color indicates the features that are too large. The method detects both local and diffuse malformations, as well as heterogeneous malformations in the same subject.

Table 1 shows the inter-subject agreement rates among neuroradiologists and with the classifier in the “blind-diagnoses” experiment. The average agreement rate among neuroradiologists is 67.6%, with a SD of 6.4%, while the classifier shows a higher agreement rate (average=73.6%, SD=7.4%). In particular, the classifier shows the highest agreement rate with the neuroradiologist not involved in the training-set selection (NR#4).

Table 2 reports the agreement rates from the ADT-diagnoses experiment. They significantly improved in comparison with the “blind-diagnoses” ones (p=0.002). More precisely, all of them increased, with the only exception of the NR#1-#4 combination, that did not change. The average agreement rate among neuroradiologists in the ADT-diagnoses is 79.3% with a SD of 4.2% and it is due to a larger agreement with the classifier suggested diagnosis (average=83.4%, SD=3.4%).

In this study, we proposed a tool that not only automatically detects the malformations of the corpus callosum from MRI images, but also provides their quantitative characterization at single-subject level. This method can be used to support physicians in the evaluation of the shape of the corpus callosum, significantly improving the inter-subject reliability. Moreover, it can be used for quantitative research studies. DD analysis results can be used in studies with a small amount of patients, where the group tests do not have sufficient statistical power. Moreover, it can be used to correlate the malformation degree to the clinical outcome, to monitor the development of the malformative pattern in pediatric patients, to detect pathology subtypes. Finally, the method has been proposed and validated on the corpus callosum, but can be extended to other brain districts.