Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Predicting Lymph Node Metastasis Through Automated and Reliable Multi-Objective Model in Head & Neck Cancer

Z Zhou1*, L Chen2, M Dohopolski3, D Sher4, J Wang5, (1) University Of Central Missouri, Warrensburg, MO, (2) UT Southwestern Medical Center, Allen, TX, (3) UT Southwestern Medical Center, Dallas, TX, (4) UT Southwestern Moncrief Cancer Ctr, Fort Worth, TX, (5) UT Southwestern Medical Center, Dallas, TX


PO-GePV-M-261 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: Accurate diagnosis of lymph node metastasis (LNM) is critical in treatment management for patients with head & neck (H&N) cancer. We aim to develop an automated and reliable multi-objective model (ARMO) for reliable LNM prediction.

Methods: Totally 129 surgical H&N cancer patients with pathology-confirmed LNs status are used and there are 130 malignant nodes and 413 benign nodes. Two hundred and fifty-six hand-crafted feature are extracted from PET and CT images, respectively. To obtain balanced outcome between sensitivity and specificity, a multi-objective model is developed to obtain Pareto-optimal model set in the training stage. In addition, to obtain calibrated confidence, which aims at maximizing the output probability with correct prediction label and minimizing the probability with incorrect prediction label, a weight that measures the relative importance of each Pareto-optimal model is calculated in training stage. Meanwhile, the individual reliability which measures how reliable of each Pareto-optimal model output is estimated in testing stage. Furthermore, an analytic evidential reasoning rule is used to combine the output of each Pareto-optimal model. Finally, the overall reliability of the model output is obtained by averaging the individual reliability to estimate the model uncertainty, which can be used to alter clinicians on model outputs with high uncertainty. Accuracy (ACC), AUC, sensitivity (SEN) and specificity (SPE) are used to measure the model performance, whilst expected calibration error (ECE) and maximum calibration error (MCE) are used to measure the confidence of calibration, where both ECE and MCE of a perfectly calibrated model are 0.

Results: The mean ± standard deviation of AUC, ACC, SEN and SPE for ARMO are 0.97±0.00, 0.93±0.00, 0.88±0.01, 0.94±0.00, respectively. Meanwhile, the ECE is 0.0085±0.0018 and MCE is 0.4483±0.0336.

Conclusion: We developed a reliable LNM prediction model, which can not only obtain promising performance, but also better calibrated confidence and uncertainty estimation.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email