Non-invasive computational immunohistochemical staining based on deep learning and multimodal imaging

In most European countries, the diagnosis of cancer is achieved by examination of haematoxylin-eosin (HE) staining by an experienced pathologist. Nevertheless, several other diagnostic approaches exist (e.g., immunohistochemical staining) which are not applied routinely for all cases due to their technical complexity, duration, and cost. Therefore, an important unmet medical need for fast, non-invasive, and label-free immunohistochemical staining based on molecular imaging without laborious sample treatment exists. This demanding challenge will be tackled in STAIN-IT using a non-invasive label-free measurement technique called multimodal imaging (e.g., the combination of coherent anti-Stokes Raman scattering, second harmonic generation, and two-photon-excited fluorescence). The multimodal images will be analysed using deep learning approaches, such as convolution neural networks (CNNs). These CNNs are utilized to mimic immunohistochemical stainings. CNNs are neural networks that learn the feature representation of the data, which is optimally suited to model a specific immunohistochemical staining. In STAIN-IT, the staining models will be developed along with the methods to quantitatively understand the nonlinear behaviour of the CNNs. With the envisioned approximation approaches for CNNs, these models no longer act as black box systems, and a quantification of tissue changes associated with the staining models can be achieved. For the very first time, STAIN-IT will develop a label-free, non-invasive, labour-inexpensive, and fast computational immunohistochemical staining, which can be easily implemented into clinical routine yielding increased diagnostic reliability and a better understanding of disease pathogenesis. A fast test of the antigen KI-67 in an intraoperative frozen section consultation situation or the use of Collagen IV as a quality control marker of tissue-engineered medicines are some of the exciting application possibilities of such staining model.