Due to major artifacts and uncalibrated Hounsfield units (HU), cone-beam computed tomography (CBCT) cannot be used readily for diagnostics and therapy planning purposes. This study addresses image-to-image translation by convolutional neural networks (CNNs) to convert CBCT to CT-like scans, comparing supervised to unsupervised training techniques, exploiting a pelvic CT/CBCT publicly available dataset. Interestingly, quantitative results were in favor of supervised against unsupervised approach showing improvements in the HU accuracy (62% vs. 50%), structural similarity index (2.5% vs. 1.1%) and peak signal-to-noise ratio (15% vs. 8%). Qualitative results conversely showcased higher anatomical artifacts in the synthetic CBCT generated by the supervised techniques. This was motivated by the higher sensitivity of the supervised training technique to the pixel-wise correspondence contained in the loss function. The unsupervised technique does not require correspondence and mitigates this drawback as it combines adversarial, cycle consistency, and identity loss functions. Overall, two main impacts qualify the paper: (a) the feasibility of CNN to generate accurate synthetic CT from CBCT images, which is fast and easy to use compared to traditional techniques applied in clinics; (b) the proposal of guidelines to drive the selection of the better training technique, which can be shifted to more general image-to-image translation.

Evidence is rapidly accumulating that multifactorial nocturnal monitoring, through the coupling of wearable devices and deep learning, may be disruptive for early diagnosis and assessment of sleep disorders. In this work, optical, differential air-pressure and acceleration signals, acquired by a chest-worn sensor, are elaborated into five somnographic-like signals, which are then used to feed a deep network. This addresses a three-fold classification problem to predict the overall signal quality (normal, corrupted), three breathing-related patterns (normal, apnea, irregular) and three sleep-related patterns (normal, snoring, noise). In order to promote explainability, the developed architecture generates additional information in the form of qualitative (saliency maps) and quantitative (confidence indices) data, which helps to improve the interpretation of the predictions. Twenty healthy subjects enrolled in this study were monitored overnight for approximately ten hours during sleep. Somnographic-like signals were manually labeled according to the three class sets to build the training dataset. Both record- and subject-wise analyses were performed to evaluate the prediction performance and the coherence of the results. The network was accurate (0.96) in distinguishing normal from corrupted signals. Breathing patterns were predicted with higher accuracy (0.93) than sleep patterns (0.76). The prediction of irregular breathing was less accurate (0.88) than that of apnea (0.97). In the sleep pattern set, the distinction between snoring (0.73) and noise events (0.61) was less effective. The confidence index associated with the prediction allowed us to elucidate ambiguous predictions better. The saliency map analysis provided useful insights to relate predictions to the input signal content. While preliminary, this work supported the recent perspective on the use of deep learning to detect particular sleep events in multiple somnographic signals, thus representing a step towards bringing the use of AI-based tools for sleep disorder detection incrementally closer to clinical translation.

Purpose: Cone beam computed tomography (CBCT) is a standard solution for in‐room image guidance for radiation therapy. It is used to evaluate and compensate for anatomopathological changes between the dose delivery plan and the fraction delivery day. CBCT is a fast and versatile solution, but it suffers from drawbacks like low contrast and requires proper calibration to derive density values. Although these limitations are even more prominent with in‐room customized CBCT systems, strategies based on deep learning have shown potential in improving image quality. As such, this article presents a method based on a convolutional neural network and a novel two‐step supervised training based on the transfer learning paradigm for shading correction in CBCT volumes with narrow field of view (FOV) acquired with an ad hoc in‐room system.Methods: We designed a U‐Net convolutional neural network, trained on axial slices of corresponding CT/CBCT couples. To improve the generalization capability of the network, we exploited two‐stage learning using two distinct data sets. At first, the network weights were trained using synthetic CBCT scans generated from a public data set, and then only the deepest layers of the network were trained again with real‐world clinical data to fine‐tune the weights. Synthetic data were generated according to real data acquisition parameters. The network takes a single grayscale volume as input and outputs the same volume with corrected shading and improved HU values.Results: Evaluation was carried out with a leave‐one‐out cross‐validation, computed on 18 unique CT/CBCT pairs from six different patients from a real‐world dataset. Comparing original CBCT to CT and improved CBCT to CT, we obtained an average improvement of 6 dB on peak signal‐to‐noise ratio (PSNR), +2% on structural similarity index measure (SSIM). The median interquartile range (IQR) Hounsfield unit (HU) difference between CBCT and CT improved from 161.37 (162.54) HU to 49.41 (66.70) HU. Region of interest (ROI)‐based HU difference was narrowed by 75% in the spongy bone (femoral head), 89% in the bladder, 85% for fat, and 83% for muscle. The improvement in contrast‐to‐noise ratio for these ROIs was about 67%.Conclusions: We demonstrated that shading correction obtaining CT‐compatible data from narrow‐FOV CBCTs acquired with a customized in‐room system is possible. Moreover, the transfer learning approach proved particularly beneficial for such a shading correction approach.

Bone segmentation and 3D reconstruction are crucial for total knee arthroplasty (TKA) surgical planning with Personalized Surgical Instruments (PSIs). Traditional semi-automatic approaches are time-consuming and operator-dependent, although they provide reliable outcomes. Moreover, the recent expansion of artificial intelligence (AI) tools towards various medical domains is transforming modern healthcare. Accordingly, this study introduces an automated AI-based pipeline to replace the current operator-based tibia and femur 3D reconstruction procedure enhancing TKA preoperative planning. Leveraging an 822 CT image dataset, a novel patch-based method and an improved segmentation label generation algorithm were coupled to a Combined Edge Loss UNet (CEL-UNet), a novel CNN architecture featuring an additional decoding branch to boost the bone boundary segmentation. Root Mean Squared Errors and Hausdorff distances compared the predicted surfaces to the reference bones showing median and interquartile values of 0.26 (0.19–0.36) mm and 0.24 (0.18–0.32) mm, and of 1.06 (0.73–2.15) mm and 1.43 (0.82–2.86) mm for the tibia and femur, respectively, outperforming previous results of our group, state-of-the-art, and UNet models. A feasibility analysis for a PSI-based surgical plan revealed sub-millimetric distance errors and sub-angular alignment uncertainties in the PSI contact areas and the two cutting planes. Finally, operational environment testing underscored the pipeline’s efficiency. More than half of the processed cases complied with the PSI prototyping requirements, reducing the overall time from 35 min to 13.1 s, while the remaining ones underwent a manual refinement step to achieve such PSI requirements, performing the procedure four to eleven times faster than the manufacturer standards. To conclude, this research advocates the need for real-world applicability and optimization of AI solutions in orthopedic surgical practice.

Objective: Recent advancements in augmented reality led to planning and navigation systems for orthopedic surgery. However little is known about mixed reality (MR) in orthopedics. Furthermore, artificial intelligence (AI) has the potential to boost the capabilities of MR by enabling automation and personalization. The purpose of this work is to assess Holoknee prototype, based on AI and MR for multimodal data visualization and surgical planning in knee osteotomy, developed to run on the HoloLens 2 headset. Methods: Two preclinical test sessions were performed with 11 participants (eight surgeons, two residents, and one medical student) executing three times six tasks, corresponding to a number of holographic data interactions and preoperative planning steps. At the end of each session, participants answered a questionnaire on user perception and usability. Results: During the second trial, the participants were faster in all tasks than in the first one, while in the third one, the time of execution decreased only for two tasks (“Patient selection” and “Scrolling through radiograph”) with respect to the second attempt, but without statistically significant difference (respectively p =0.14 and p =0.13, p <0.05). All subjects strongly agreed that MR can be used effectively for surgical training, whereas 10 (90.9%) strongly agreed that it can be used effectively for preoperative planning. Six (54.5%) agreed and two of them (18.2%) strongly agreed that it can be used effectively for intraoperative guidance. Discussion/Conclusion: In this work, we presented Holoknee, the first holistic application of AI and MR for surgical planning for knee osteotomy. It reported promising results on its potential translation to surgical training, preoperative planning, and surgical guidance.

Eye tracking techniques based on deep learning are rapidly spreading in a wide variety of application fields. With this study, we want to exploit the potentiality of eye tracking techniques in ocular proton therapy (OPT) applications. We implemented a fully automatic approach based on two-stage convolutional neural networks (CNNs): the first stage roughly identifies the eye position and the second one performs a fine iris and pupil detection. We selected 707 video frames recorded during clinical operations during OPT treatments performed at our institute. 650 frames were used for training and 57 for a blind test. The estimations of iris and pupil were evaluated against the manual labelled contours delineated by a clinical operator. For iris and pupil predictions, Dice coefficient (median = 0.94 and 0.97), Szymkiewicz–Simpson coefficient (median = 0.97 and 0.98), Intersection over Union coefficient (median = 0.88 and 0.94) and Hausdorff distance (median = 11.6 and 5.0 (pixels)) were quantified. Iris and pupil regions were found to be comparable to the manually labelled ground truths. Our proposed framework could provide an automatic approach to quantitatively evaluating pupil and iris misalignments, and it could be used as an additional support tool for clinical activity, without impacting in any way with the consolidated routine.