The lack of interpretability of deep learning reduces understanding of what happens when a network does not work as expected and hinders its use in critical fields like medicine, which require transparency of decisions. For example, a healthy vs pathological classification model should rely on radiological signs and not on some training dataset biases. Several post-hoc models have been proposed to explain the decision of a trained network. However, they are very seldom used to enforce interpretability during training and none in accordance with the classification. In this paper, we propose a new weakly supervised method for both interpretable healthy vs pathological classification and anomaly detection. A new loss function is added to a standard classification model to constrain each voxel of healthy images to drive the network decision towards the healthy class according to gradient-based attributions. This constraint reveals pathological structures for patient images, allowing their unsupervised segmentation. Moreover, we advocate both theoretically and experimentally, that constrained training with the simple Gradient attribution is similar to constraints with the heavier Expected Gradient, consequently reducing the computational cost. We also propose a combination of attributions during the constrained training making the model robust to the attribution choice at inference. Our proposition was evaluated on two brain pathologies: tumors and multiple sclerosis. This new constraint provides a more relevant classification, with a more pathology-driven decision. For anomaly detection, the proposed method outperforms state-of-the-art especially on difficult multiple sclerosis lesions segmentation task with a 15 points Dice improvement.
1 March 2021 in Biochimica et Biophysica Acta (BBA) - Molecular Cell Research
Because our civilization has relied on pesticides to fight weeds, insects, and diseases since antiquity, the use of these chemicals has become natural and exclusive. Unfortunately, the use of pesticides has progressively had alarming effects on water quality, biodiversity, and human health. This paper proposes to improve farming practices by replacing pesticides with a laser-based robotic approach. This study focused on the neutralization of aphids, as they are among the most harmful pests for crops and complex to control. With the help of deep learning, we developed a mobile robot that spans crop rows, locates aphids, and neutralizes them with laser beams. We have built a prototype with the sole purpose of validating the localization-neutralization loop on a single seedling row. The experiments performed in our laboratory demonstrate the feasibility of detecting different lines of aphids (50% detected at 3 cm/s) and of neutralizing them (90% mortality) without impacting the growth of their host plants. The results are encouraging since aphids are one of the most challenging crop pests to eradicate. However, enhancements in detection and mainly in targeting are necessary to be useful in a real farming context. Moreover, robustness regarding field conditions should be evaluated.
/sup 18/F-labeled NaF, also called [/sup 18/F] fluoride ion, Positron Emission Tomography (NaF PET) is a specific imaging modality of bone activity and allows bone tumor detection. In this paper, we propose a fast-automated method for the location of anatomical structures by special planes and seeds in whole body NaF PET images. This step is crucial in registration and segmentation processes such as region growing. Our method proceeds in two steps: first, it delineates anatomical objects such as bladder, skull, spine and legs with two horizontal planes (bottom and top); secondly, it determines labeled seeds in these objects. Processes are based on the analysis of the Slice Energy (SE) and the computation of the Center of Gravity (CG) of all transverse slices. This method was applied to seeds initialization in order to segment the whole skeleton. This segmentation could be helpful for bone tumors detection and bone uptake quantification. Fast and satisfying results are obtained. Robustness of our method was positively tested on a set of seven NaF PET images.