Despite all the progress made in the development of new therapeutic methods, cancer remains one of the most serious health hazards and most frequent causes of death worldwide. For optimal treatment of patients, early and accurate histopathological diagnosis of tumors and their respective subtypes is of utmost relevance. Although some diagnostic tools are available for this purpose in clinical routine, a considerable degree of uncertainty remains in some cases.
The DIAMANT project focuses on lung cancer, the most common cancer in men and the second most common in women worldwide. The major subtypes are adenocarcinoma (ADC) and squamous cell carcinoma (SqCC), which together account for ~70% of all lung cancer diagnoses. The distinction between these subtypes is very important, as some therapies are promising in one type but contraindicated in the other. However, in poorly differentiated tumors, typing is often not possible based on morphology alone, but requires additional immunohistochemical (IHC) staining with up to four markers. Less extensive workup in these cases leads to decreased diagnostic accuracy, with the risk of delayed or even incorrect diagnosis.
To address this problem in the context of tumor classification, the use of imaging matrix-assisted laser desorption/ionization mass spectrometry (MALDI IMS) has been proposed. MALDI IMS is an analytical method for spatially resolved analysis of proteins, metabolites, lipids, and other molecules in biological tissue samples. A major advantage is its applicability to formalin-fixed, paraffin-embedded tissue sections, the vast majority of which are used in routine histopathology. In recent years, promising results have been obtained with regard to MALDI IMS-based classification of various cancer types. However, the method has not yet been established in clinical routine. One reason for this is the limited spatial resolution of MALDI IMS (approximately 50 μm), which does not allow molecular imaging at the cellular level.
Therefore, in DIAMANT, molecular information from MALDI IMS is combined with detailed anatomical information from digital microscopy images (Digital Image Analysis, DIA). Using an integrated analysis of the data from both complementary modalities, a classification model will be developed that is significantly more accurate than existing models based on only one of the two modalities. The deep learning methodology used in the project is particularly well suited to extract meaningful information from the complex, high-dimensional raw data from both modalities. Using an application-specific network architecture and specially adapted regularization methods, pathological, biochemical and mass spectrometric a priori knowledge is integrated into the classification model. The required amount of training data is thus reduced and the robustness of the method and its generalizability to new clinical test data are increased.
