Background: Cortical lesions are highly inconspicuous on magnetic resonance imaging (MRI). Double inversion recovery (DIR) has a higher sensitivity than conventional clinical sequences (i.e. T1, T2, FLAIR) but is difficult to acquire, leading to overseen cortical lesions in clinical care and clinical trials. Objective: To evaluate the usability of artificially generated DIR (aDIR) images for cortical lesion detection compared to conventionally acquired DIR (cDIR). Methods: The dataset consisted of 3D-T1 and 2D-proton density (PD) T2 images of 73 patients (49RR, 20SP, 4PP) at 1.5 T. Using a 4:1 train:test-ratio, a fully convolutional neural network was trained to predict 3D-aDIR from 3D-T1 and 2D-PD/T2 images. Randomized blind scoring of the test set was used to determine detection reliability, precision and recall. Results: A total of 626 vs 696 cortical lesions were detected on 15 aDIR vs cDIR images (intraclass correlation coefficient (ICC) = 0.92). Compared to cDIR, precision and recall were 0.84 ± 0.06 and 0.76 ± 0.09, respectively. The frontal and temporal lobes showed the largest differences in discernibility. Conclusion: Cortical lesions can be detected with good reliability on artificial DIR. The technique has potential to broaden the availability of DIR in clinical care and provides the opportunity of ex post facto implementation of cortical lesions imaging in existing clinical trial data.

In retinoblastoma, accurate segmentation of ocular structure and tumor tissue is important when working towards personalized treatment. This retrospective study serves to evaluate the performance of multi-view convolutional neural networks (MV-CNNs) for automated eye and tumor segmentation on MRI in retinoblastoma patients. Forty retinoblastoma and 20 healthy-eyes from 30 patients were included in a train/test (N = 29 retinoblastoma-, 17 healthy-eyes) and independent validation (N = 11 retinoblastoma-, 3 healthy-eyes) set. Imaging was done using 3.0 T Fast Imaging Employing Steady-state Acquisition (FIESTA), T2-weighted and contrast-enhanced T1-weighted sequences. Sclera, vitreous humour, lens, retinal detachment and tumor were manually delineated on FIESTA images to serve as a reference standard. Volumetric and spatial performance were assessed by calculating intra-class correlation (ICC) and dice similarity coefficient (DSC). Additionally, the effects of multi-scale, sequences and data augmentation were explored. Optimal performance was obtained by using a three-level pyramid MV-CNN with FIESTA, T2 and T1c sequences and data augmentation. Eye and tumor volumetric ICC were 0.997 and 0.996, respectively. Median [Interquartile range] DSC for eye, sclera, vitreous, lens, retinal detachment and tumor were 0.965 [0.950–0.975], 0.847 [0.782–0.893], 0.975 [0.930–0.986], 0.909 [0.847–0.951], 0.828 [0.458–0.962] and 0.914 [0.852–0.958], respectively. MV-CNN can be used to obtain accurate ocular structure and tumor segmentations in retinoblastoma.

Depending on the clinical situation, different combinations of lymph node (LN) levels define the elective LN target volume in head-and-neck cancer (HNC) radiotherapy. The accurate auto-contouring of individual LN levels could reduce the burden and variability of manual segmentation and be used regardless of the primary tumor location. We evaluated three deep learning approaches for the segmenting individual LN levels I–V, which were manually contoured on CT scans from 70 HNC patients. The networks were trained and evaluated using five-fold cross-validation and ensemble learning for 60 patients with (1) 3D patch-based UNets, (2) multi-view (MV) voxel classification networks and (3) sequential UNet+MV. The performances were evaluated using Dice similarity coefficients (DSC) for automated and manual segmentations for individual levels, and the planning target volumes were extrapolated from the combined levels I–V and II–IV, both for the cross-validation and for an independent test set of 10 patients. The median DSC were 0.80, 0.66 and 0.82 for UNet, MV and UNet+MV, respectively. Overall, UNet+MV significantly (p < 0.0001) outperformed other arrangements and yielded DSC = 0.87, 0.85, 0.86, 0.82, 0.77, 0.77 for the combined and individual level I–V structures, respectively. Both PTVs were also significantly (p < 0.0001) more accurate with UNet+MV, with DSC = 0.91 and 0.90, respectively. The accurate segmentation of individual LN levels I–V can be achieved using an ensemble of UNets. UNet+MV can further refine this result.