ABSTRACT
Rectal cancer is a common cancer and leading cause of cancer death. The majority of rectal cancer patients
are diagnosed with stage II or III locally advanced disease. Neoadjuvant chemoradiotherapy followed by total
mesorectal excision is the standard of care for treating these patients. However, the response to neoadjuvant
therapy is highly heterogeneous. While approximately 20% patients achieve a pathologic complete response
and long-term disease control, over 30% patients will develop distant metastasis despite multimodality
treatment. Because total mesorectal excision is associated with significant morbidity and poor quality of life,
organ preservation (i.e., watchful waiting without surgery) is recommended for patients who are deemed to
achieve a complete response. On the other hand, for patients at high risk of recurrence, total neoadjuvant
therapy with induction chemotherapy is increasingly used for early eradication of micrometastases to improve
cure rate. The successful implementation of these two highly promising treatment strategies (organ
preservation and total neoadjuvant therapy) hinges on the precise knowledge of (1) which patients will have a
pathologic complete response; and (2) which patients will develop distant metastasis. Unfortunately, current
clinical tools are rather crude and do not allow for accurate outcome prediction on an individual basis, leading
to over-treatment in some patients and under-treatment in others. There is a pressing unmet need for reliable
prognostic and predictive biomarkers to guide personalized treatment of rectal cancer.
To address this unmet need, we adopt a rational approach for prognosis prediction by developing radiomic and
deep learning models that are informed and guided by established knowledge of the tumor pathobiology.
Additionally, we propose novel approaches to analyze serial images for predicting pathologic response to
neoadjuvant therapy. Further, by leveraging the complementary value of imaging and blood-based biomarkers,
we will construct integrative models to improve outcome prediction. To establish clinical validity, we will use a
large retrospective dataset for model training and conduct rigorous prospective validation. If successful, this
project will lead to risk-adaptive and response-driven personalized treatment strategies, which may ultimately
improve outcomes for patients with rectal cancer.