Click here to

Session: Image-Guided Surgery and Interventions [Return to Session]

Deep Learning for Near Real-Time Image-Guided Focal Ablation

B Anderson*, B Rigaud, Y Lin, E Lin, G Cazoulat, E Koay, A Jones, B Odisio, K Brock, UT MD Anderson Cancer Center, Houston, TX


MO-A-TRACK 6-2 (Monday, 7/26/2021) 10:30 AM - 11:30 AM [Eastern Time (GMT-4)]

Purpose: Harness deep-learning based segmentation and registration capabilities combined with biomechanical model-based registration to train an outcome prediction algorithm for liver focal ablation outcomes.

Methods: 129 patients with 204 lesions who underwent ablation therapy were retrospectively evaluated, including follow-up status of the ablated lesion. Liver, and disease segmentations were defined on pre-treatment and post-treatment CT images using a clinically validated AI-based algorithm. Lesions were separated based on patient MRN into a training/validation cohort, with 29 locally progressing and 24 non-progressing sites with > 6 month follow-up in the validation cohort. Image features from the pre-treatment CT were extracted via convolutions and fed into fully-connected layers for outcome prediction. Tuned hyper-parameters included dense layers, dense connections, convolutional layers, dropout, transition blocks, number of filters, unique loss functions (cosine/categorical cross entropy), and the images presented. We trained and evaluated the model using both rigid and deformable image registration (DIR), including liver segmentations, and/or disease segmentations based on validation area under the curve (AUC).

Results: Qualitative assessment of our AI-based segmentation models showed immediate clinical usability or < 10 seconds editing in 96% (48/50) of liver, and 100% (24/24) of disease segmentations. AUC values were maximized when presenting the model with DIR images and the disease contour (AUC=0.75), and when presenting the model with rigid images, liver, and disease contours (AUC=0.74).

Conclusion: Our fully automated outcome prediction model can provide the clinician with a likelihood of future tumor progression based on intra-procedural imaging in a clinically actionable timeframe (< 1 minute), providing feedback on ablation efficacy.

Funding Support, Disclosures, and Conflict of Interest: Funding was provided by the Society of Interventional Radiology Allied Scientist Grant, and the Dr. John J Kopchick Fellowship.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email