• 43007 Tarragona, Spain

Title of Project : Determination of molecular subtypes of BC using multimodal radiological imaging                                            
Student Name: Mariia Donskova      

To develop a new methodology to automate the classification of the different molecular subtypes of BC (i.e., Luminal A, Luminal B, Triple-negative and HR-/HER2-) from the analysis of radiological images (e.g., mammograms, MRI and ultrasound (traditional and elastography)), by the following steps: (1) segmentation of the tumour by means of a DL model (e.g., GAN); (2) obtaining fractal, morphological and kinetic textural characteristics of the segmented tumour regions in the previous step, which will be calculated using extractors of characteristics adapted to the visual patterns present in the digital images, as well as the features extracted by the DL model; (3) studying the correlation between the tumour shape and margins predicted and molecular subtypes of BC, which will be compared to the pioneer clinical studies; (4) developing an explainable DL model based on CNN for trustworthy classification of molecular subtypes of BC, (5) investigating the correlation between the molecular subtypes and the risk to BC relapse, and (6) finally validating the results of the proposed system in the hospital of Sant Joan de Reus (associated partner), with whom the candidate is regularly interacting, via screening radiological images for the entire population of Catalunya between 2005 and 2015. We will use our previous work for breast tumour segmentation and shape classification in mammograms using generative adversarial and convolutional neural networks as a baseline of this thesis to achieve more accurate and reliable results and to use more different radiological images. Our hypothesis will be corroborated on INbreast (115 cases), DDSM (1168 cases) and private in-house (from HUSJR) (300 cases) datasets.

As a result, we will obtain 1) a powerful biomarker based on explainable DL models including visual, fractal, morphological and kinetic textural characteristics represented by a unique feature vector for each subtype, leaving behind the obsolete handcrafted features produced by traditional descriptors, which did not allow such specialization, 2) the correlation of the texture patterns that appear in the images and define the internal zone of the tumour, with the different prognostic factors of each subtype (estrogen and progesterone receptors, Ki67, PVL,...), 3) a classification model based on an explainable DL model based on the correlation of the tumour appearance in radiological images and tumour subtypes (especially with Luminal A, Luminal B and Triple Negative carcinomas).