Ensemble machine learning model architecture for lesion detection

Inventors

Bonakdar Sakhi, Omid • Esquinas Fernandez, Pedro Luis • PALMA, Giovanni John Jacques • Dufort, Paul

Assignees

Guerbet SA

Abstract

A lesion detection ensemble machine learning model architecture comprising a plurality of trained machine learning (ML) computer models is provided. A first decoder of a lesion detection ML model processes a medical image input to generate a first lesion mapping prediction. A second decoder of the lesion detection ML model processes the medical image input to generate a second lesion mapping prediction. Combinational logic combines the first and second lesion mapping predictions to generate a combined prediction. Final lesion mapping output logic generates a final lesion prediction based on the combined lesion mapping prediction. The final lesion mapping output logic outputs the final lesion prediction for further downstream computing operations. The first decoder is trained with a first loss function that is configured to counterbalance a training of the second decoder that is trained using a second loss function different from the first loss function.

Core Innovation

The invention relates to a lesion detection ensemble machine learning model architecture comprising a plurality of trained machine learning computer models. The architecture processes a medical image input using a first decoder and a second decoder to generate first and second lesion mapping prediction outputs identifying graphical elements corresponding to lesions.

The lesion detection machine learning computer model combines the first lesion mapping prediction output and the second lesion mapping prediction output using combinational logic to generate a combined lesion mapping prediction output. Based on the combined lesion mapping prediction output, final lesion mapping output logic generates a final lesion prediction output, which is output for further downstream computing operations.

The first decoder is trained with a first loss function configured to counterbalance a training of the second decoder, and the second decoder is trained using a second loss function different from the first loss function. In additional described implementations, the training includes loss-function behaviors tied to counterbalancing false-negative lesion detection and false-positive lesion detection, and optionally includes a third loss function configured to force the first lesion mapping prediction output and the second lesion mapping prediction output to be consistent with each other.

The pipeline can further include anatomical-structure masking using a mask generation machine learning computer model to generate a liver mask and apply it to produce a masked portion of the medical image input corresponding to an anatomical structure of interest. The architecture can then produce lesion mapping predictions for masked regions and also produce unmasked lesion mapping predictions, with a final lesion prediction derived from combining the masked and unmasked lesion mapping prediction output, including an average combination rule in one described implementation.

Claims Coverage

The document provides four independent claims centered on a two-decoder lesion-detection ensemble with combined lesion mapping and a final lesion prediction output. The inventive features include the ensemble workflow, the use of combinational logic and final lesion mapping output logic, and training counterbalancing by using different loss functions for the first and second decoders. Additional dependent claims refine loss behaviors, optionally add mask-generation for an anatomical structure, and extend final prediction aggregation including averaging.

Across the independent claims, the core claim coverage is the two-decoder ensemble workflow that produces separate lesion mapping prediction outputs, combines them via combinational logic into a combined lesion mapping prediction output, and derives a final lesion prediction output using final lesion mapping output logic. The independent claims further specify training counterbalancing by assigning different loss functions to the first and second decoders. The dependent claims additionally cover loss semantics, an optional third loss for consistency between decoder outputs, optional anatomical-structure mask generation, and final prediction aggregation that can incorporate unmasked predictions and averaging.

Stated Advantages

Documented Applications