Automatic hippocampal subfield segmentation in ultra-high resolution MRI using Convolutional Neural Network

Jacobsen, Nina ; Nøhr, Anne Krogh ; Hansen, Bolette Dybkjær

4. term

Biomedical Engineering and Informatics, Master

2017

2017-06-22

Hippocampus’ subfelter er vigtige i forskningen, fordi ændringer her kan forbindes med neurodegenerative sygdomme. Bedre adgang til MR-skanninger med ultrahøjt magnetfelt (7T) har gjort detaljeret segmentering af disse områder mulig. Der findes automatiske metoder, men de er tidskrævende og har ofte svært ved at segmentere de mindste subfelter. Konvolutionelle neurale netværk (CNN) har derimod vist hurtig og nøjagtig segmentering af hjernestrukturer. Vi udviklede derfor DeepHSS, en CNN-baseret metode til automatisk segmentering af hippocampus’ subfelter, og undersøgte effekten af et lille datasæt. DeepHSS omfatter en forbehandlingspipeline efterfulgt af et 12-lags 3D-CNN med to veje, trænet med tæt (dense) træning. Træningssættet bestod af 7T TSE MR-billeder fra 15 personer med manuelle markeringer af seks hippocampale subfelter samt entorhinal cortex. Metoden blev testet på 10 personer. Vi sammenlignede de automatiske segmenteringer med de manuelle ved hjælp af Dice Similarity Coefficient (DSC), et standardmål for overlap hvor 1,0 er perfekt. DeepHSS leverede hurtige segmenteringer med en gennemsnitlig DSC på de mærkede strukturer på 0,83 ± 0,03. De højeste gennemsnitlige DSC-værdier blev opnået for subiculum (0,80 ± 0,05) og dentate gyrus (0,77 ± 0,05), mens den laveste var for entorhinal cortex (0,49 ± 0,14). Nøjagtigheden steg med antallet af træningspersoner. Samlet viser DeepHSS, at CNN’er effektivt kan segmentere hippocampale subfelter selv med et relativt lille datasæt, med resultater der er sammenlignelige med ASHS.

Hippocampal subfields are a focus of research because changes in these areas are linked to neurodegenerative diseases. Greater access to ultra–high-field MRI (7T) has made detailed segmentation of these regions feasible. Existing automatic methods are often slow and struggle with the smallest subfields, whereas convolutional neural networks (CNNs) have shown fast, accurate segmentation in the brain. We developed DeepHSS, a CNN-based method for automatic segmentation of hippocampal subfields, and examined how a small dataset affects performance. DeepHSS includes a preprocessing pipeline followed by a 12-layer, two‑pathway 3D CNN trained using dense training. The training set comprised 7T TSE MRI from 15 participants with manual labels for six hippocampal subfields and the entorhinal cortex. The method was tested on 10 participants. We compared the automatic segmentations to manual labels using the Dice Similarity Coefficient (DSC), a standard measure of overlap where 1.0 indicates a perfect match. DeepHSS produced segmentations quickly, achieving an average DSC across labeled structures of 0.83 ± 0.03. The highest average DSCs were for the subiculum (0.80 ± 0.05) and dentate gyrus (0.77 ± 0.05), and the lowest for the entorhinal cortex (0.49 ± 0.14). Accuracy increased with the number of training subjects. Overall, DeepHSS shows that CNNs can efficiently segment hippocampal subfields even with a relatively small dataset, with results comparable to ASHS.

[This abstract was generated with the help of AI]

CNN ; convolutional neural network ; segmentation ; MRI ; hippocampal subfields ; ultra-high field ; deep learning

Download
View record in AAU Student Projects