System and method for analyzing a map of a biological object with optimized longitudinal compatibility

A computer-implemented method for providing a trained main machine learning algorithm configured for outputting a change mask M_changes when receiving as input longitudinal maps of a part of a biological object, wherein the change mask M_changes is configured for highlighting or showing longitudinal changes between the longitudinal maps, wherein the main ML algorithm comprises a second ML algorithm, called N_long, and optionally a first ML algorithm, called N_cross, wherein N_cross is configured for segmenting a map received as input and for providing as output a segmentation mask for the map received as input, wherein N_long is configured for detecting longitudinal changes between two longitudinal maps received as input, the computer-implemented method comprising:

• acquiring a first N_long input training dataset comprising, for each training biological object of a set of training biological objects, at least one map representing said part at a first time point and at least one corresponding map representing said part at a second time point different from the first time point, so that, for each training biological object, its part is represented by at least two maps, being longitudinal maps, representing the part at said first time point and said second time point, respectively;
• acquiring a second N_long input training dataset comprising, for each training biological object of said set of training biological objects, for each map representing the part at the first time point, a first training segmentation mask M_T_1 representing said part of the training biological object, wherein said first training segmentation mask M_T_1 has been obtained from said map representing said part at the first time point for said training biological object, and, for each corresponding map representing the part at the second time point, a corresponding second training segmentation mask M_T_2 representing said part, wherein said second training segmentation mask M_T_2 has been obtained from said map representing said part at the second time point different from the first time point;
• acquiring an N_long output training dataset comprising, for each training biological object of said set of training biological objects, and for each couple formed by a first segmentation mask and corresponding second segmentation mask thereof, a related training change mask M_T_changes representing longitudinal changes that took place between said map representing the training biological object part at the first time point and its corresponding map representing said part of the training biological object at the second time point;
• for each of said training biological objects, training N_long for said detection of longitudinal changes based on said first N_long input training dataset, said second N_long input training dataset, and said N_long output training dataset defined for the concerned training biological object;
• providing the trained main ML algorithm via an output interface;
• wherein, for each training biological object, the main ML algorithm is trained to optimize or maximize a compatibility between the first training segmentation mask M_T_1, the second training segmentation mask M_T_2, and M_changes.

A computer-implemented method for providing a trained main machine learning algorithm configured for outputting a segmentation of at least one of longitudinal maps of a part of a biological object when receiving as input said longitudinal maps, wherein said main ML algorithm comprises a first ML algorithm, called N_cross, and optionally a second ML algorithm, called N_long, wherein N_cross is configured for segmenting a map received as input and for providing as output a segmentation mask for the received map, wherein N_long is configured for detecting longitudinal changes between two longitudinal maps received as input by outputting a change mask M_changes representing said longitudinal changes, wherein training the main ML algorithm comprises training at least N_cross, the computer-implemented method comprising:

• acquiring an N_cross input training dataset comprising, for each training biological object of a set of training biological objects, at least one fist map representing said part at a first time point, and at least one second map representing said part at a second time point different from the first time point, so that, for each training biological object, its part is represented by at least two maps, being longitudinal maps, representing the part at two different time points, being said first time point and said second time point;
• for each training biological object, acquiring an N_cross output training dataset comprising, for each of said training biological objects, and for each of the maps representing said part at the first time point, a segmentation mask, called prior ground truth mask GT_prior, and/or, for each of the maps representing said part at the second time point, another segmentation mask called current ground truth mask GT_current;
• training N_cross for said segmentation of longitudinal images based on the N_cross input training dataset and the N_cross output training dataset;
• providing the trained main ML algorithm via an output interface;
• wherein the main ML algorithm is trained to optimize or maximize a compatibility between at least one of M1 or M2 and M_T_changes, wherein, for each training biological object, and, for a concerned training biological object, for each couple formed by a first map and a second map, there is a corresponding M_T_changes that is a change mask representing longitudinal changes between said first map and said second map of the couple.