Britt Suelmann

UMC Utrecht

In short, Britt Suelmann's research entails the following:

Artificial-intelligence based prediction of absence of vital retroperitoneal residual disease and teratoma in metastatic germ cell tumors after platinum-based chemotherapy

Testicular germ cell tumors (TGCTs) are a rare form of cancer, accounting for approximately 1% of all cancers in men, with the highest incidence among those aged 15 to 40 years. In the Netherlands, 500-900 cases are diagnosed annually, and about one-third of these involve metastatic disease at diagnosis. With the advent of cisplatin-based chemotherapy, the survival rate for metastatic TGCTs has improved significantly, exceeding 90%. However, about 33% of patients are left with residual retroperitoneal masses after chemotherapy. These masses are composed of teratoma (45%), necrosis or fibrosis (45%), and vital tumor cells (10%).

Currently, all patients with residual retroperitoneal masses undergo surgical resection, known as retroperitoneal lymph node dissection (RPLND), to remove these masses. This approach is necessary because imaging techniques like CT and MRI cannot reliably differentiate between being (necrosis/fibrosis) and malignant (teratoma/vital tumor) tissue. As a result, nearly half of the patients undergo potentially unnecessary, complex, and risky surgeries for masses that do not require
treatment. These surgeries carry significant risks of complications, including infection, bleeding, and long-term health issues. 

The inability of current diagnostic tools to distinguish between types of residual masses underscores the need for more precise methods. Clinical factors, such as tumor markers, primary tumor histopathology, and changes in mass size during chemotherapy, have been explored as predictors but lack sufficient accuracy for routine use.

The aim of this research is to develop an artificial intelligence (AI)-based model to predict the absence of vital tumor cells or teratoma in residual masses, thereby sparing patients from unnecessary surgery. This approach leverages contrast-enhanced (CE) CT scans and combines imaging data with clinical information to improve diagnostic accuracy. Two AI methodologies are being investigated: deep radiomics and weakly supervised learning. Deep radiomics involves analyzing CT-scan features with neural networks and correlating them with clinical data using Random Forest classifiers. Weakly supervised learning directly predicts the absence of residual disease from CT scans using neural networks, with explanations provided by interpretability tools like Shapley additive explanations. 

The study will utilize data from 500 patients treated at two Dutch referral centers between 2011 and 2023. This dataset includes longitudinal CT scans, clinical parameters such as tumor markers, chemotherapy regimens, and pathology results. AI models will be trained and evaluated using nested cross-validation and Receiver Operating Characteristic (ROC) analysis to optimize performance and minimize false negatives. The ultimate goal is to reduce unnecessary surgeries by accurately identifying patients with benign residual masses.

The potential impact of this AI-based method is significant. By sparing patients from extensive and harmful surgeries, it can reduce physical and emotional burdens, improve long-term quality of life, and lower healthcare costs. Furthermore, the approach could be expanded to other cancers where distinguishing residual disease after therapy remains challenging. This innovation not only aims to transform the management of TGCTs but also represents a broader step forward in precision
oncology, leveraging AI to enhance diagnostic accuracy and patient outcomes.