About model $M_{R_{WHM},R_{tum}}$

[Devil-Ideal]

Hi , I have read your paper, the work is awesome. But I'm not quite sure how the model $M{R{WHM},R_{tum}}$ was trained. It was trained on the real FLAIR volumes from the SABRE dataset, and the BRATS tumor volumes were used to train the synthetic model, leaving the training labels empty. Since the labels are empty, how can it be trained?

[virginiafdez]

Hello,

I guess you refer to the part about training segmentation nn-UNets to test the data. In that case, you are training to segment WMH, not tumours. Supposing you do not have WMH masks, the idea is that you can still try to make the nnU-Net model see tumour images where the tumour area - which otherwise the model will be labelling as WMH as seen from the other trainings - is not labelled as WMH (the labels are blank). You train this model along with SABRE images and labels pairs which do have WMH. This, supposedly, is attempting to reduce the number of false positives (tumour pixels labelled as WMH). However, as you do have WMH lesions in the BRATS dataset, the model wrongly associates the presence of tumour to an absence of WMH lesions.

Hope this clarifies it.

Best regards,

Virginia

---

From: Devil-Ideal @.> Sent: 28 November 2023 13:31 To: virginiafdez/brainSPADE3D_rel @.> Cc: Subscribed @.***> Subject: [virginiafdez/brainSPADE3Drel] About model $M{R{WHM},R{tum}}$ (Issue #1)

Hi , I have read your paper, the work is awesome. But I'm not quite sure how the model $M{R{WHM},R_{tum}}$ was trained. It was trained on the real FLAIR volumes from the SABRE dataset, and the BRATS tumor volumes were used to train the synthetic model, leaving the training labels empty. Since the labels are empty, how can it be trained?

— Reply to this email directly, view it on GitHubhttps://github.com/virginiafdez/brainSPADE3D_rel/issues/1, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AOVQGV5TF2XXIV2Q73O5WM3YGXRT5AVCNFSM6AAAAAA75YSLMOVHI2DSMVQWIX3LMV43ASLTON2WKOZSGAYTINBUHE2TQMQ. You are receiving this because you are subscribed to this thread.Message ID: @.***>