MENU

Click here to

√ó

Are you sure ?

Yes, do it No, cancel

Detecting Pathological Complete Response in Esophageal Cancer After Neoadjuvant Therapy Based On Survival-Weighted Deep Learning: A Pilot Study

S Cheng1*, W Yap2, E Tu3, (1) Taiwan AI Labs, ,,(2) Chang Gung Memorial Hospital, ,,(3) Taiwan Ai Labs,

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
According to a recent meta-analysis, 24% to 32% of patients with esophageal cancer reach a pathological complete response (pCR) after neoadjuvant radiochemotherapy (nRCT). For complete responders to nRCT, watchful-waiting approach instead of esophagectomy is currently under investigation. However, current image modalities seem to be insufficiently accurate to identify complete responders. Here, we introduce a deep supervised learning method leveraging a survival-weighted loss function to predict pCR while focusing on accuracy of subgroup with shorter progression-free survival (PFS).

Methods:
All node-positive esophageal squamous cell carcinoma patients treated with nRCT followed by surgery between January 2014 and December 2017 are reviewed. Patients are categorized into pCR (ypT0/Tis ypN0) group and non-pCR group. A dual-path DenseNet model is trained on two channels of pixel data (planning CT and total dose map) for pCR prediction. The input of one path is GTV/CTV block in 3D shape for extracting pT features and the other input is whole slices in 2.5D shape for acquiring pN information. The loss function of cross-entropy is weighted by the value of customized PFS function w(s) (Supplementary_Figure_1).

Results:
80 patients are included, of them 23 had pCR and 57 had non-pCR. Patient and tumor characteristics are summarized (Supplementary_Table_1). 43 patients (53.75%) with shorter PFS than average (21.05 months) are stratified into high-risk subgroup. The dataset is randomly split into training-set (15 pCR + 41 non-pCR; 31 high-risk) and testing-set (8 pCR + 16 non-pCR; 12 high-risk). When survival-weighting is applied, the accuracy is 0.9167 (95% CI: 0.8979-0.9355) of high-risk subgroup versus 0.750 (0.7312-0.7688) of the others. When survival-weighting is not applied, the accuracy is 0.8333 (0.8079-0.8587) of high-risk subgroup versus 0.8333 (0.8101-0.8565) of the others.

Conclusion:
The accuracy of pCR prediction in shorter PFS subgroup is improved by survival-weighted learning without significant decrease in accuracy for the whole cohort.

Download ePoster [PDF]

Keywords

ROC Analysis, Decision Theory, Dose Response

Taxonomy

TH- Dataset Analysis/Biomathematics: Machine learning techniques

Contact Email