功能性磁振造影解碼之全腦特徵選擇方法

Abstract

腦部解碼旨在解釋神經活動以理解大腦功能。本論文探討了各種全腦特徵選擇方法及先進深度學習技術在功能性磁共振成像（fMRI）數據中的應用，旨在提高腦部解碼能力。我們提出了一種新的兩階段特徵選擇方法，將大腦體積分成非重疊的立方體，有效識別高貢獻的體素。這種方法顯著提高了來自個體內數據的音高解碼準確性。

利用自監督學習，我們運用對比學習在人類連接組計劃（HCP）中的多個任務fMRI資料集上，以學習可轉移的特徵。我們的方法在解碼各種心理狀態方面顯示出顯著改進，相較於現有方法，尤其在數據有限的情況下，我們的方法與稀疏字典方法比較下具有競爭力。通過在預訓練階段使用多樣的任務fMRI資料，我們提升了下游任務的性能。

此外，我們探討了學習到的特徵在類似條件下的可轉移性，例如解碼音樂類型。我們的實驗顯示，任務相關的預訓練在數據有限的情況下更加高效和有效。我們發現，古典音樂和嘻哈音樂更易區分，這可能是由於音樂片段中的歌詞影響了語言處理區域。雖然我們的方法受益於多個受試者數據的聚合，傳統特徵選擇方法在某些任務中仍具有優勢。總的來說，本論文強調了使用全腦fMRI數據和自監督學習在推進神經影像分析和解碼複雜神經活動方面的潛力。

Brain decoding seeks to interpret neural activity to identify mental states. This thesis explores various whole-brain feature selection methods and the application of advanced deep learning techniques to enhance brain decoding capabilities using functional magnetic resonance imaging (fMRI) data. We propose a novel two-stage feature selection method that divides brain volumes into non-overlapping cubes, effectively identifying high-contribution voxels. This approach significantly improves decoding accuracy for musical pitch information from within-subject data.

Leveraging self-supervised learning, we apply contrastive learning to multiple task fMRI datasets from the Human Connectome Project (HCP) to learn transferable features. Our method demonstrates substantial improvements in decoding various mental states compared to existing approaches and proves competitive with sparse dictionary methods, particularly when data is limited. By using diverse task fMRI data during pretraining, we enhance downstream task performance.

Additionally, we investigate the transferability of learned features to similar conditions, such as decoding musical genres. Our experiments reveal that task-related pretraining can be more efficient and effective, especially with limited data. We find that classical and hip-hop music are more distinguishable, likely due to the presence of lyrics affecting language processing regions. While our method benefits from aggregating data from multiple subjects, traditional feature selection methods still excel in certain tasks. Overall, this thesis underscores the potential of using whole-brain fMRI data and self-supervised learning to advance neuroimaging analysis and decode complex neural activities.