Automation of Wilms' tumor segmentation by artificial intelligence.

BACKGROUND: 3D reconstruction of Wilms' tumor provides several advantages but are not systematically performed because manual segmentation is extremely time-consuming. The objective of our study was to develop an artificial intelligence tool to automate the segmentation of tumors and kidneys in children. METHODS: A manual segmentation was carried out by two experts on 14 CT scans. Then, the segmentation of Wilms' tumor and neoplastic kidney was automatically performed using the CNN U-Net and the same CNN U-Net trained according to the OV2ASSION method. The time saving for the expert was estimated depending on the number of sections automatically segmented. RESULTS: When segmentations were performed manually by two experts, the inter-individual variability resulted in a Dice index of 0.95 for tumor and 0.87 for kidney. Fully automatic segmentation with the CNN U-Net yielded a poor Dice index of 0.69 for Wilms' tumor and 0.27 for kidney. With the OV2ASSION method, the Dice index varied depending on the number of manually segmented sections. For the segmentation of the Wilms' tumor and neoplastic kidney, it varied respectively from 0.97 to 0.94 for a gap of 1 (2 out of 3 sections performed manually) to 0.94 and 0.86 for a gap of 10 (1 section out of 6 performed manually). CONCLUSION: Fully automated segmentation remains a challenge in the field of medical image processing. Although it is possible to use already developed neural networks, such as U-Net, we found that the results obtained were not satisfactory for segmentation of neoplastic kidneys or Wilms' tumors in children. We developed an innovative CNN U-Net training method that makes it possible to segment the kidney and its tumor with the same precision as an expert while reducing their intervention time by 80%.

A Deep Learning-Based Automated Framework for Subpeak Designation on Intracranial Pressure Signals.

The intracranial pressure (ICP) signal, as monitored on patients in intensive care units, contains pulses of cardiac origin, where P1 and P2 subpeaks can often be observed. When calculable, the ratio of their relative amplitudes is an indicator of the patient's cerebral compliance. This characterization is particularly informative for the overall state of the cerebrospinal system. The aim of this study is to develop and assess the performances of a deep learning-based pipeline for P2/P1 ratio computation that only takes a raw ICP signal as an input. The output P2/P1 ratio signal can be discontinuous since P1 and P2 subpeaks are not always visible. The proposed pipeline performs four tasks, namely (i) heartbeat-induced pulse detection, (ii) pulse selection, (iii) P1 and P2 designation, and (iv) signal smoothing and outlier removal. For tasks (i) and (ii), the performance of a recurrent neural network is compared to that of a convolutional neural network. The final algorithm is evaluated on a 4344-pulse testing dataset sampled from 10 patient recordings. Pulse selection is achieved with an area under the curve of 0.90, whereas the subpeak designation algorithm identifies pulses with a P2/P1 ratio > 1 with 97.3% accuracy. Although it still needs to be evaluated on a larger number of labeled recordings, our automated P2/P1 ratio calculation framework appears to be a promising tool that can be easily embedded into bedside monitoring devices.

Risk prediction in surgery using case-based reasoning and agent-based modelization.

Managing the risks arising from the actions and conditions of the various elements that make up an operating room is a major concern during a surgical procedure. One of the main challenges is to define alert thresholds in a non-deterministic context where unpredictable adverse events occur. In response to this problematic, this paper presents an architecture that couples a Multi-Agent System (MAS) with Case-Based Reasoning (CBR). The possibility of emulating a large number of situations thanks to MAS, combined with analytical data management thanks to CBR, is an original and efficient way of determining thresholds that are not defined a priori. We also compared different similarity calculation methods (Retrieve phase of CBR). The results presented in this article show that our model can manage alert thresholds in an environment that manages data as disparate as infectious agents, patient's vitals and human fatigue. In addition, they reveal that the thresholds proposed by the system are more efficient than the predefined ones. These results tend to prove that our simulator is an effective alert generator. Nevertheless, the context remains a simulation mode that we would like to enrich with real data from, for example, monitoring sensors (bracelet for human fatigue, monitoring, etc).

3D reconstruction of Wilms' tumor and kidneys in children: Variability, usefulness and constraints.

INTRODUCTION: Wilms' tumor (WT) is the most common type of malignant kidney tumor in children. Three-dimensional reconstructions can be performed pre-operatively to help surgeons in the planning phase. OBJECTIVES: The main objective of this study was to determine the variability of WT segmentation and 3D reconstruction. The secondary objectives were to assess the usefulness of these 3D reconstructions in the surgical planning phase and in the selection of patients for nephron-sparing surgery (NSS). METHODS: 14 scans from 12 patients were manually or semi-automatically segmented by 2 teams using 3D Slicer software. Inter-individual variability of 3D reconstructions was measured based on the Dice index. The utility of 3D reconstructions for the surgical planning was evaluated by 4 pediatric surgeons using a 5-point Likert scale. The possibility of undertaking NSS was evaluated according to the criteria defined in the Umbrella SIOP-RTSG 2016 protocol. RESULTS: Segmentation of the WT, healthy kidney, pathological kidney, arterial and venous vascularization could be performed for all of the patients in this study. Urinary cavities segmentation could only be performed for 5 out of 14 scans that had a delayed acquisition phase. The mean time required to carry out these segmentations was 8.6 h [3-15 h]. The mean Dice index for all of the scans was good (mean: 0.87; range [0.83-0.91]). Considering each anatomical structure, the Dice index was very good for the WT (mean: 0.95; range [0.91-0.97]) and the healthy kidney (mean: 0.95; range [0.93-0.96]), good for the pathological kidney (mean: 0.87; range [0.69-0.96]) and arterial vascularization (mean: 0.84; range [0.74-0.91]). The Dice index was lower than 0.8 for venous vascularization only (mean: 0.77; range [0.58-0.86]). All the surgeons who were interviewed agreed that the 3D reconstructions were realistic representations and useful for the surgical planning phase. The images reconstructed in 3D allowed most of the criteria defined by the Umbrella SIOP-RTSG 2016 protocol to be evaluated regarding the selection of patients who could benefit from NSS. CONCLUSION: The inter-individual variability of 3D reconstructions of WT is acceptable. Three-dimensional representation appears to assist surgeons with the surgical planning phase by allowing them to better anticipate the operative risks. 3D reconstructions can also be an additional tool to better select patients for NSS. However, the manual or semi-automatic method used is very time-consuming, making it difficult for a routinely use. Developing techniques to automate this segmentation process, therefore, appears to be essential if surgeons and radiologists are to use it in daily practice.

Fusion of multiple segmentations of medical images using OV2ASSION and Deep Learning methods: Application to CT-Scans for tumoral kidney.

Nephroblastoma is the most common kidney tumour in children. Its diagnosis is based on imagery. In the SAIAD project, we have designed a platform for optimizing the segmentation of deformed kidney and tumour with a small dataset, using Artificial Intelligence methods. These patient's structures segmented by separate tools and processes must then be fused to obtain a unique numerical 3D representation. However, when aggregating these structures into a final segmentation, conflicting pixels may appear. These conflicts can be solved by IA techniques. This paper presents a synthesis of our segmentation contribution in the SAIAD project and a new fusion method. The segmentation method uses the FCN-8s network with the OV2ASSION training method, which allows segmentation by patient and overcomes the limited dataset. This new fusion method combines the segmentations of the previously performed structures, using a simple and efficient network combined with the OV2ASSION training method as well, in order to manage eventual conflicting pixels. These segmentation and fusion methods were evaluated on pathological kidney and tumour structures of 14 patients affected by nephroblastoma, included in the final dataset of the SAIAD project. They are compared with other methods adapted from the literature. The results demonstrate the effectiveness of our training method coupled with the FCN-8s network in the segmentation process with more patients, and in the case of the fusion process, its effectiveness coupled with a common network, in resolving the conflicting pixels and its ability to improve the resulting segmentations.

Conflict management in the fusion of complementary segmentations of deformed kidneys and nephroblastoma.

The fusion of multiple segmentations aims to improve their accuracy in order to make them exploitable. However, conflicts may appear. In this paper, two conflict-management models are proposed for the fusion of complementary segmentations. This conflict-management and fusion procedure, integrated into the SAIAD project, carries out the fusion of deformed kidneys and nephroblastoma using the combination of six independent methods. These methods are based on different criteria, like the adjacent segmented slices, the variation of information, the Dice, the neighbouring labels, the pixel intensity by scanner images, and the fully connected CRFs. The performances of our fusion models was evaluated on 139 scans for three patients with nephroblastoma, and the results demonstrate its effectiveness and the improvement of the resulting segmentations.

Development of a 4D numerical chest phantom with customizable breathing.

Respiratory movement information is useful for radiation therapy, and is generally obtained using 4D scanners (4DCT). In the interest of patient safety, reducing the use of 4DCT could be a significant step in reducing radiation exposure, the effects of which are not well documented. The authors propose a customized 4D numerical phantom representing the organ contours. Firstly, breathing movement can be simulated and customized according to the patient's anthroporadiametric data. Using learning sets constituted by 4D scanners, artificial neural networks can be trained to interpolate the lung contours corresponding to an unknown patient, and then to simulate its respiration. Lung movement during the breathing cycle is modeled by predicting the lung contours at any respiratory phases. The interpolation is validated comparing the obtained lung contours with 4DCT via Dice coefficient. Secondly, a preliminary study of cardiac and œsophageal motion is also presented to demonstrate the flexibility of this approach. The application may simulate the position and volume of the lungs, the œsophagus and the heart at every phase of the respiratory cycle with a good accuracy: the validation of the lung modeling gives a Dice index greater than 0.93 with 4DCT over a breath cycle.