Many studies presented at ESMO Congress 2022 show the potential of artificial intelligence to enhance personalised oncology
Returning to our labs and practices after a busy and bustling ESMO Congress 2022, we are all excited about the myriad research advances, novel treatment strategies, technological developments and emerging biomarkers in precision cancer medicine that showcased on the ground in Paris.
Featuring throughout the scientific programme par excellence were several studies evidencing how artificial intelligence (AI) is increasingly enhancing personalized cancer care by potentiating diagnostic capability and improving the molecular characterization of tumors and their microenvironment toward ultimately improving patient outcomes.
Enhancing cancer patient stratification
Over recent years, technological advances in oncology have led to the development of digital pathology techniques for the acquisition, management and interpretation of pathology information using whole slide imaging and AI-based solutions.
The utilization of digitalized methodologies is expanding our understanding of the tumour microenvironment and enabling researchers to study and extract information beyond the human eye. As importantly, these approaches promise to better guide patient stratification and selection as well as support the multidisciplinary identification of optimal treatment strategies based on patient profiles.
Results of a study presented at ESMO Congress 2022, directed by researchers at the Cleveland Clinic, Cleveland, U.S., show that AI-powered pathology can help guide pathologists to better stratify patients at diagnosis and thus possibly avoid treatment with toxic therapy (Abstract 593P). By training machine learning models using digitalized hematoxylin and eosin-stained whole slide images, the researchers have evidenced that AI-powered pathology can classify cells and tissue regions in the high-grade serous carcinoma tumour microenvironment and that nuclear morphology associates with patient outcomes.
Data also reveal an association between cancer cell nuclear morphological diversity, aneuploidy and patient prognosis, and support the use of digital pathology as a dynamic, image-based environment that enables the identification of objective, quantitative biomarkers that may help pathologists to better stratify patients at diagnosis and avoid potentially toxic treatments, which is of paramount importance as we collectively seek to identify ‘kinder’ treatment strategies matched to the molecular specificities of individual tumors.
Machine learning and drug response prediction
Considerable advances in deep learning are also paving the way in the development of more accurate tools for therapy response in patients. As an example, results of a study led by colleagues at the Stanford University School of Medicine (USA), also presented at the ESMO Congress, report on the predictive performance an AI-based platform developed to enhance routine pre-treatment histopathology specimens to predict neoadjuvant chemotherapy response in patients with muscle-invasive bladder cancer (Abstract 1773P).
As indicated by the authors, pathological complete response to cisplatin-based neoadjuvant chemotherapy in this patient population is only around 30%, illuminating the need to identify robust biomarkers to guide patient selection more effectively as well as minimize therapy-related morbidity and improve treatment outcomes.
Using clinical information and pre-treatment hematoxylin and eosin-stained histopathology specimens, the deep learning method was trained to extract quantitative, cell-type-specific features including nuclear geometry, cellular spatial arrangement, and tissue heterogeneity, that were then correlated with cancer-specific survival using a multivariable model.
The machine learning prediction model demonstrated a significant predictive performance and shows promise in identifying those patients who would be most likely to respond to neoadjuvant chemotherapy and thus aiding clinical decision-making.
Deep learning, deeper insights
Novel anti-HER2 antibody drug conjugates have shown anti-tumour activity against HER2-low breast cancer, but a subset of tumours classified as HER2-negative by current immunohistochemistry testing methods may also express this protein.
Colleagues at the Memorial Sloan Kettering Cancer Center, New York, U.S., have now shown that deep learning methods applied to hematoxylin and eosin-stained whole slide images can more accurately detect tumours expressing HER2 (Abstract 93P). Data show that their AI system can increase diagnostic confidence by distinguishing between breast cancers lacking any HER2 protein and mRNA and HER2-low tumors.
As indicated by the investigators, this AI-driven approach will need to be further validated in additional cohorts of patients treated with new anti-HER2 antibody-drug conjugates to evaluate its potential in more precisely guiding treatment decisions as well as possibly informing the design of new biomarkers for drug development.
Evaluating deep learning for predicting HER2-positive status from routinely obtained haematoxylin and eosin-stained whole slide images, a study led by researchers at the Berlin Institute for the Foundations of Leaning and Data – BIFOLD, Germany, put their model to the test using large datasets from patients enrolled in three neoadjuvant clinical studies, and also used selective prediction analysis to identify a subset of patients matched to the aggregated machine learning features (Abstract 68MO).
Results show the high performance of this deep learning method in predicting HER2 status, and that the generalization of this approach to a held-out clinical study as well as analysis in subsets of patients further validate its value. Considering these data, this AI-driven model could potentially be used to support the selection of targeted therapies in breast cancer patients.
AI-based prediction of disease relapse in colon cancer
Representing an unmet clinical need, up to 40% of patients with localized colon cancer relapse despite optimal initial therapy, and while circulating tumour DNA has emerged as a novel prognostic marker in this setting, the accuracy in predicting disease recurrence is somewhat limited. Results of a single-center retrospective observational study spearheaded by investigators at the INCLIVA Biomedical Research Institute, Valencia, Spain, show promise in more accurately anticipating relapse using machine learning models combining radiomics and deep learning features extracted from computed tomography images (Abstract 340P).
Using real-world clinical data and CT imaging from 60 patients, their predictive model showed a significant increase in accuracy compared with clinical variables alone: 95% versus 55%. Findings evidence that the extraction of radiomics and deep features from CT images and their combination with AI algorithms provide important insights into the prediction of cancer recurrence and patient outcomes and could help to identify and develop new and more precise prognostic imaging panels.
The future ahead
While still largely in their infancy, AI-driven approaches continue to step up in improving cancer imaging analysis, predicting treatment outcomes in individual patients (risk factors, possible effects of therapy, cancer drug response, and survival), increasing the accuracy and speed of diagnosis, and supporting patient selection for clinical trials.
Much work still needs to be done to translate their vast potential into real-world and clinically meaningful applications, but this ‘smart’, wide-ranging branch of computer science is gaining pace in expanding the armory of precision medicine in oncology.
Considering the future integration of machine learning approaches as an additional tool in the multidisciplinary management of our patients, it is a definite OUI from me!
