Oral Presentation 2025 Joint Meeting of the COSA ASM and IPOS Congress

Improving chemotherapy toxicity prediction in colon cancer patients using AI-CT derived 3D body composition of the entire L1–L5 lumbar spine (126396)

Ke Cao 1 , Josephine Yeung 1 , Matthew YK Wei 1 2 , CheukShan Choi 1 , Margaret Lee 3 , Lincoln J Lim 1 4 , Yasser Arafat 1 2 , Paul N Baird 5 , Justin MC Yeung 1 2
  1. Department of Surgery, Western Precinct, The University of Melbourne, Melbourne, VIC, Australia
  2. Department of Colorectal Surgery, Western Health, Melbourne, VIC, Australia
  3. Department of Oncology, Western Health, Melbourne, VIC, Australia
  4. Department of Radiology, Western Health, Melbourne, VIC, Australia
  5. Department of Surgery, University of Melbourne, Melbourne, VIC, Australia

Objectives

Chemotherapy dose-limiting toxicities (DLT) pose a significant challenge in successful colon cancer treatment. Body composition analysis may enable tailored interventions thereby supporting the mitigation of chemotherapy toxic effects. This study aimed to evaluate and compare the effectiveness of using three-dimensional (3D) computed tomography (CT)–derived body composition measures from the entire lumbar spine levels (L1-L5) versus a single vertebral level (L3), the current gold standard, in predicting chemotherapy DLT in colon cancer patients.

 

Sample and setting

This retrospective study included 184 non-metastatic colon cancer patients who received adjuvant chemotherapy at Western Health, Melbourne.

 

Procedures

DLT were defined as any dose reduction or discontinuation due to chemotherapy toxicity. Using artificial intelligence auto-segmentation, 3D body composition measurements, including the volume (in cm³) of muscle, adipose tissues and bone, were obtained from patients’ L1-5 levels on CT imaging. The effectiveness of using patients’ 3D body composition measurements from the L3 level, as well as incorporating data from the entire L1–L5 region (including L3), in predicting DLT was evaluated using machine learning techniques.

 

Results

Of the 184 patients, 112 (60.9%) experienced DLT. Neuropathy was the most common toxicity (49/112, 43.8%), followed by diarrhoea (35.7%) and nausea/vomiting (33%). Patients with DLT had significantly lower muscle (p = 0.005) and bone volume (p = 0.04) in the lumbar spine region (L1–L5) compared to those with no DLT. The machine learning model incorporating L1-L5 data and patient clinical data achieved a higher predictive performance (accuracy=0.76), that outperformed prediction using a single L3 level (accuracy=0.66).

 

Conclusion and clinical implications

Evaluating a patient’s body composition allowed prediction of chemotherapy toxicities for colon cancer. Incorporating fully automated body composition analysis of CT slices from the entire lumbar region offers promising performance in early identification of high-risk individuals, with the ultimate aim of improving patient’s quality of life.