QS107 - Classification of Endoscopic Rectal Adenocarcinoma Treatment Response Using Deep Learning Models

Sunday, June 2, 2024

8:42 AM – 8:45 AM ET, USA

Location: Room 327

Quick Shot Presenter(s)

Seth Felder, MD

Associate Member, Department of Gastrointestinal Oncology
Moffitt Cancer Center
Moffitt Cancer Center
Tampa, FL, United States

All relevant financial disclosures have been managed and will be listed below.
This individual has no financial relationships with ineligible companies.

Author(s)

ZW

Zhouping Wei, PhD

Applied Research Scientist, Machine Learning
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.
SA

Shatha Abudalou, n/a

Graduate/Ph.D. Research Student
Moffitt Cancer Center; Unversity of South Florida
Tampa, Florida, United States

Disclosure information not submitted.
RP

Raj Mohan Paspulati, MD

Senior Member, Diagnostic Imaging
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.
JC

James Costello, MD, PhD

Associate Member, Diagnostic Imaging
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.
HV

Harshna Vadvala, MD

Assistant Member, Diagnostic Imaging
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.
SD

Sophie Dessureault, MD, PhD

Senior Member
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.
AS

Amalia Stefanou, MD FACS FASCRS

Assistant Member
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.
Julian Sanchez, MD (he/him/his)

Senior Member, Department of Gastrointestinal Oncology
Moffitt Cancer Center
Moffitt Cancer Center
Tampa, Florida, United States

All relevant financial disclosures have been managed and will be listed below.
This individual has no financial relationships with ineligible companies.

Senior Author(s)

YB

Yoganand Balagurunathan, PhD

Assistant Member, Machine Learning
Moffitt Cancer Center
Tampa, Florida, United States

Disclosure information not submitted.

Purpose/Background: Patient selection for rectal organ preservation following receipt of neoadjuvant treatment(s) is suboptimal and associated with a consistent rate of local tumor regrowth. Deep learning models analyzing post-neoadjuvant treatment endoscopic digital images may be an ancillary decision aid to improve the predictive accuracy of mucosally based tumor response extent, influencing patient selection for mesorectal excision or a Watch and Wait (WW) strategy.

Methods/Interventions: The digital re-staging endoscopic optical 2D image snapshots from a retrospective cohort of 39 patients (172 unique 2D-slices) with locally advanced rectal cancer patients following neoadjuvant treatment(s) between 2017-2022 were identified. A pre-trained convolutional neural network or deep learning (DL) network (VGG V16, visual geometry) architecture using ImageNet database then shallow trained (varied 10% to 50%) with a small proportion of the rectal cancer data set. The DL networks predictive performance on the test data (not used in training) was evaluated using sensitivity, specificity, and accuracy. Visual class activation maps (Grad-CAM) were compared with the clinical insights to discern the DL networks ability to identify spatial patterns.

Results/Outcomes: Endoscopic re-staging images from 39 patients were used to obtain 172 optical 2D slices in the dataset. 9 patients (31 slices) were designated as a sustained clinical complete response (>12 months) or underwent mesorectal excision and pathologically confirmed a ypT0pN0, 84 slices designated near complete response and 57 slices designated incomplete response. The VGG 16 DL network accuracy of discrimination (true tumor/no tumor) was 76%, sensitivity 46%, and specificity 89%.

Conclusion/Discussion: A DL model analyzing re-staged rectal cancer endoscopic images may decrease clinician subjectivity interpreting endoscopic tumor response extent to improve patient selection for organ preservation. The model demonstrated modest accuracy in distinguishing endoscopic incomplete response (near complete + incomplete response) from complete responders, however, the dataset is small limiting its predictive ability. Grad-CAM maps providing spatial visual patterns may complement standard endoscopic interpretation of tumor response extent. Increasing the image set and combining clinical findings with biomarkers along with MRI radiomic data may allow clinicians to best predict the degree of tumor response in the context of rectal organ preservation. Machine learning technologies applied to routinely obtained endoscopic-radiographic images may become a reliable clinical tool when selecting rectal cancer patients for organ preservation.