The CNN was trained on control T1w images (n = 1049) from the HCP Dataset—Diffusion MRI 3T 1200 Subjects (S1200) Release (Glasser et al., 2013; Van Essen et al., 2013) downloaded from the brainlife.io platform (Avesani et al., 2019)². As detailed in the PreFreeSurfer pipeline from HCP Minimal Preprocessing Pipelines (Glasser et al., 2013), for each subject, the T1w images acquired with native 0.7-mm isotropic resolution were defaced (Milchenko and Marcus, 2013), aligned to MNI152 template space (rigid-body transformation with 6 degrees of freedom), and corrected for readout distortions (van der Kouwe et al., 2008). The preprocessed images were further resampled to 1.25-mm isotropic resolution to match the resolution of the HCP DWI data. Importantly, the downsampling was also a prerequisite for further segmentation of T1w images with FreeSurfer software.

The performance of the trained CNN was tested on the T1w MRI images of PWA (n = 9) and controls (n = 8) from the CHIASM dataset (Puzniak et al., in revision, see footnote 1) downloaded from the brainlife.io platform (Avesani et al., 2019)³. As preprocessing steps, for each subject, T1w images acquired with native resolution of 0.9 mm were defaced, aligned to Anterior Commissure—Posterior Commissure (ACPC) space, and downsampled to 1-mm isotropic voxel (in order to support FreeSurfer segmentation).

The X-mask_manual were defined in PWA (n = 9; CHIASM albinism) and controls (n = 8; CHIASM controls) from the CHIASM dataset, and 10 HCP test-controls were excluded from CNN training. Specifically, the delineation was performed by a trained researcher in all the T1w image slices with chiasmal presence, as according to the guidelines detailed in (Puzniak et al., in revision, see footnote 1). The X-mask_manual were deemed a ground truth and consequently used as reference for the quality assessment of other masks (Figure 2A).

The X-mask_{atlas–initial} were extracted from the existing atlas-based segmentation of the HCP T1w images (Van Essen et al., 2013) processed according to the HCP FreeSurfer pipeline (Glasser et al., 2013) using FreeSurfer v5.2 (Fischl, 2012). Although such atlas-based masks were successfully used in previous studies aiming to accelerate brain segmentation using CNNs (Fedorov et al., 2017a, b; McClure et al., 2019), our comparison of X-mask_{atlas–initial} with X-mask_manual revealed a significantly lower quality of the former (see Results), thereby making them (as expected) a suboptimal choice as training data.

Although X-mask_{atlas–initial} were found to be of insufficient quality for training, we observed that their shortcomings can be mitigated by incorporating information about voxel intensities in the mask delineation process. This allowed us to formulate the following seven-step algorithm generating a corrected mask, X-mask_{atlas–corrected} from the X-mask_{atlas–initial}:

The quantitative comparison (see Computational methods) of outcome for the X-mask_{atlas–corrected} with X-mask_manual demonstrated a significantly improved quality of the former (see Results) in comparison to X-mask_{atlas–initial}. Given this validation, the correction procedure was subsequently performed for all of the HCP X-mask_{atlas–initial}, and the resulting X-mask_{atlas–corrected} were used as targets for supervised training of the CNN.

The developed CNN used a 3D version (Çiçek et al., 2016) of the U-Net architecture (Ronneberger et al., 2015). Although the 3D version involves a higher computational load which may limit the upper resolution of processed images, the inclusion of additional dimension was shown to be of significant benefit to the segmentation (Chen et al., 2019; Mlynarski et al., 2020). Another reason for using the U-Net architecture was its reported robustness to jagged boundary-localized errors (Heller et al., 2018), which is a helpful feature in case of training on automatically generated masks.

Specifically, the network consists of analysis (encoding) and synthesis (decoding) paths. The analysis path contains four layers, each containing a standard U-net block [two 3 × 3 × 3 × convolutions followed by batch normalization and rectified linear unit (ReLu)] and subsequent 2 × 2 × 2 max pooling (stride of 2). For each subsequent step in the analysis path, the number of feature maps derived from input was doubled in each layer. The synthesis path consists of 2 × 2 × 2 upconvolution followed by a U-net block. Importantly, each decoding layer receives concatenated feature maps from a previous decoding layer and corresponding encoding layer, which allows for preservation of both low- and high-level features. Finally, in the last layer, the two output feature maps (background and target class, here optic chiasm) are being normalized with a voxelwise softmax function. The total number of parameters is 2,206,482.

Prior to being fed into the network, the training images and target X-mask_{atlas–corrected} were subjected to the following data augmentation procedure [performed using the TorchIO package (Pérez-García et al., 2021)], respectively:

Performing these steps on the training data allowed us to enhance CNN’s generalization capabilities without resorting to explicit regularization, which has been demonstrated to limit the model’s capacity (Hernández-García and König, 2018). This was particularly important considering the misalignment of HCP training images (stored in MNI152 space) and CHIASM testing images (aligned to ACPC space).

The HCP grayscale T1w images (n = 1049) and corresponding X-mask_{atlas–corrected} were divided into training (n = 932, 87.5%; HCP training—controls), validation (n = 107, 10%; HCP validation—controls), and testing (n = 10, 2.5%) subgroups augmented and subsequently fed in batches of 2 to U-Net CNN using the Dice similarity coefficient (see Computational methods) loss function and Adam optimizer (Kingma and Ba, 2017) for the purpose of weight updating. The training was performed five times, using different combinations of hyperparameters, specifically the maximal numbers of epochs (13, 15, 30, 40, 100) and learning rates (respectively, 0.0025, 0.0030, 0.0025, 0.0015, and 0.0005). The resulting weights of trained networks were saved and are provided (see Availability of data and material).

The trained CNN returned a grayscale image of the input’s size (160 × 160 × 160), where each voxel’s intensity depicted the probability of belonging to the optic chiasm (ranging from 0 to 1). This output was turned into a binary optic chiasm mask by thresholding the image to an empirically selected value (here we tested a range of 0.25, 0.50, 0.75, and 1.00 thresholds) and selecting the biggest cluster of nonzero elements present in the image. The quality of final X-mask_CNN generated by the tested range of training hyperparameters and threshold values were evaluated against X-mask_manual for both HCP test-controls and CHIASM controls datasets (see Supplementary Material), with the best-reported performance achieved for 30 epochs, and a learning rate equal to 0.0025 at a threshold of 1.

In order to measure the quality of optic chiasm masks, we employed the Dice similarity coefficient [DSC; (Dice, 1945; Sørensen, 1948)] statistic, which describes the amount of overlap between two masks, in our case. The DSC ranges from 0 (lack of overlap) to 1 (perfect overlap of identical shapes). Specifically, we calculated the value of DSC between the ground-truth X-mask_manual and the candidate X-mask_{‘candidate’}, where the latter has been previously limited only to axial slices where X-mask_manual was present (the excessive voxels were cropped). For brevity, the value of DSC calculated between X-mask_manual and candidate X-mask_{‘candidate’} is further being denoted to as DSC_{manual_vs_‘candidate’}, and in case of group-level results, statistics is presented as mean ± standard error of mean (SEM).

The obtained DSC values, grouped with respect to compared candidate mask group (X-mask_initial, X-mask_{atlas–corrected}, and X-mask_CNN) and participant group (HCP test-controls, CHIASM controls, and CHIASM albinism), were subjected to statistical testing. All samples were tested for normal distribution using the test by D’Agostino and Pearson (D’Agostino and Pearson, 1973), and for all but one (X-mask_{atlas–initial} : HCP test-controls) the null hypothesis of coming from normal distribution could not be rejected. Accordingly, in case of comparison of two normally distributed samples, we used two-tailed, two-sampled t-test at an alpha level of 5%; otherwise, we used the Wilcoxon rank-sum test, which tests the null hypothesis that two samples are drawn from the same distribution. Finally, we controlled for the familywise error (FWE) rate by applying Bonferroni’s correction to all calculated p-values.

In order to evaluate any potential discrepancy in the X-mask quality obtained for PWA and controls, we classified the obtained DSC_{manual_vs_CNN} using C-support vector classification (C-SVC) model with polynomial kernel (Platt, 1999; Chang and Lin, 2011). The measure of interclass discrepancy was subsequently quantified using well-established machine learning classification metrics, specifically

where TP, TN, FP, and FN are, respectively, true positives, true negatives, false positives, and false negatives. Importantly, it should be noted that the classifier has been trained and evaluated on the same data, which is a clearly forbidden practice in case of evaluating a classifiers’ performance. Our goal was, however, to quantitatively express the intergroup differences in DSC_{manual_vs_CNN} and the overlap in data points, which is why we decided on such an approach. In line with that purpose, we used the support vector classification model which attempts to maximize the margin around the decision boundary.

An important step in the validation of a diagnostic tool is the demonstration of its sensitivity to meaningful individualized biomarkers of the disease. Unfortunately, the evidence for such specific anatomical biomarkers of chiasmal malformations is missing. Although previous studies (Schmitz et al., 2003; von dem Hagen et al., 2005) provided a list of candidate features distinguishing normal from abnormal chiasms (width of chiasm, optic nerves and optic tracts, angle between optic tracts), none of these features were investigated in the context of individualized diagnostics. This lack of literature on the anatomy-based detection of chiasmal malformations was also the primary reason for the choice of our method here, which aimed at the investigation of the generalized applicability of CNNs for the purpose of detection of chiasmal malformations.

While the missing knowledge on specific relevant anatomical biomarkers of chiasmal misrouting can be retrieved by the identification of features driving the correct diagnostics, this process requires the CNN to be interpretable, i.e., to grant insight into identifying the input features driving the outcome decision. This property is, however, not generally available for all the CNNs but is rather dependent on their task. Specifically, in the case of described segmentation CNN (as chosen in this study to avoid dependence on the scarce data on chiasmal malformations), there are strong limits of the possibilities for visualization of the inference process. This limitation does not apply to the classification CNNs, where several interpretation techniques (Montavon et al., 2018) such as Grad-CAM (Selvaraju et al., 2017) can be implemented. These, however, require extensive datasets. Additionally, as this type of network uses both normal and abnormal data for training, it allows for extensive validation of features that drive the CNN’s decision that will lead to understand their placing in the normal–abnormal spectrum. In summary, we note that while the employed approach provides the evidence that certain anatomical biomarkers of chiasmal malformations exist, the next step should involve their identification with a classifying CNN trained on larger datasets comprising both normal and abnormal data.

Although our study’s design enabled us to take advantage of the massive HCP dataset and train the CNN on much bigger samples than typically used, the rarity of albinism (and other cases of congenital malformations of the chiasm) severely limited the size of the training sample. For the same reasons, we were limited to testing data from only one site, which prevented us from investigation of impact of scanner and data acquisition protocol on the method’s outcome. Moreover, this limited our estimates of the accuracy and robustness of the presented method.

Although the HCP dataset is well-known for setting standards in MRI data quality assessment, it does not contain any clinical information pertaining to the participants’ visual system evaluations. Furthermore, the training dataset might include participants with retinal and/or optic nerve disorders in proportions corresponding to their representation in the real world. This prevalence is, however, not expected to influence the outcome of CNN training, but even if so, this would cause underestimation of our method’s performance, rather than overestimation.

Similarly, although the CHIASM datasets provides findings of ophthalmologic examination of included participants, it does not provide information about the types of albinism represented in the dataset (e.g., oculo-cutaneous, ocular albinism).

Finally, the quality of automatically generated training labels is inferior to the ones created manually. Although we provide evidence that this does not impact the general outcome of the study, we acknowledge the use of manually defined labels to be the optimal approach.

Due to the models’ complexity and high dependence on the underlying data, the Deep Learning modeling approach is mainly driven by empirical, rather than theoretical, evidence. Consequently, despite the choice of employing an established 3D U-Net architecture, which was reported to perform well in similar task, we cannot rule out that other architectures would not provide better results. Similarly during CNN training, although we tried several combinations of hyperparameters and reported the ones yielding the best results, it is nearly impossible that we found the globally optimal configuration of the CNN network’s parameters.