The challenges facing deep learning-based catheter localization for ultrasound guided high-dose-rate prostate brachytherapy.

BACKGROUND: Automated catheter localization for ultrasound guided high-dose-rate prostate brachytherapy faces challenges relating to imaging noise and artifacts. To date, catheter reconstruction during the clinical procedure is performed manually. Deep learning has been successfully applied to a wide variety of complex tasks and has the potential to tackle the unique challenges associated with multiple catheter localization on ultrasound. Such a task is well suited for automation, with the potential to improve productivity and reliability. PURPOSE: We developed a deep learning model for automated catheter reconstruction and investigated potential factors influencing model performance. The model was designed to integrate into a clinical workflow, with a proposed reconstruction confidence metric to aid in planner verification. METHODS: Datasets from 242 patients treated from 2016 to 2020 were collected retrospectively. The anonymized dataset comprises 31,000 transverse images reconstructed from 3D sagittal ultrasound acquisitions and 3500 implanted catheters manually localized by the planner. Each catheter was retrospectively ranked based on the severity of imaging artifacts affecting reconstruction difficulty. The U-NET deep learning architecture was trained to localize implanted catheters on transverse images. A fivefold cross-validation method was used, allowing for evaluation over the entire dataset. The postprocessing software combined the predictions with patient-specific implant information to reconstructed catheters in 3D space, uniquely matched to the implanted grid positions. A reconstruction confidence metric was calculated based on the number and probability of localized predictions per catheter. For each patient, deep learning prediction and postprocessing reconstruction were completed in under 2 min on a nonperformance PC. RESULTS: Overall, 80% of catheter reconstructions were accurate, within 2 mm along 90% of the length. The catheter tip was often not detected and required extrapolation during reconstruction. The reconstruction accuracy was 89% for the easiest catheter ranking and decreased to 13% for the highest difficulty ranking, when the aid of live ultrasound would have been recommended. Even when limited to the easiest ranked catheters, the reconstruction accuracy decreased at distal grid positions, down to 50%. Individual implantation style was found to influence the frequency of severe artifacts, slightly impacting the model accuracy. A reconstruction confidence metric identified the difficult catheters, removed the observed individual variation, and increased the overall accuracy to 91% while excluding 27% of the reconstructions. CONCLUSIONS: The deep learning model localized implanted catheters over a large clinical dataset, with overall promising results. The model faced challenges due to ultrasound artifacts and image degradation distal to the probe, underlining the continued importance of maintaining image quality and minimizing artifacts. A potential workflow for integration into the clinical procedure was demonstrated, including the use of a confidence metric to predict low accuracy reconstructions. Comparison between models evaluated on different datasets should also consider underlying differences, such as the frequency and severity of imaging artifacts.

Machine Learning to Identify Gene Interactions from High-Throughput Mutant Crosses.

Advances in molecular genetics through high-throughput gene mutagenesis and genetic crossing have enabled gene interaction mapping across whole genomes. Detecting gene interactions in even small microbial genomes relies on measuring growth phenotypes in thousands of crossed strains followed by statistical analysis to compare single and double mutants. The preferred computational approach is to use a multiplicative model that factors phenotype scores of single gene mutants to identify gene interactions in double mutants. Here we present how machine learning models that consider the characteristics of the phenotypic data improve on the classical multiplicative model. Importantly, machine learning improves the selection of cutoff values to identify gene interactions from phenotypic scores.

A Gaussian process-based definition reveals new and bona fide genetic interactions compared to a multiplicative model in the Gram-negative Escherichia coli.

MOTIVATION: A digenic genetic interaction (GI) is observed when mutations in two genes within the same organism yield a phenotype that is different from the expected, given each mutation's individual effects. While multiplicative scoring is widely applied to define GIs, revealing underlying gene functions, it remains unclear if it is the most suitable choice for scoring GIs in Escherichia coli. Here, we assess many different definitions, including the multiplicative model, for mapping functional links between genes and pathways in E.coli. RESULTS: Using our published E.coli GI datasets, we show computationally that a machine learning Gaussian process (GP)-based definition better identifies functional associations among genes than a multiplicative model, which we have experimentally confirmed on a set of gene pairs. Overall, the GP definition improves the detection of GIs, biological reasoning of epistatic connectivity, as well as the quality of GI maps in E.coli, and, potentially, other microbes. AVAILABILITY AND IMPLEMENTATION: The source code and parameters used to generate the machine learning models in WEKA software were provided in the Supplementary information. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Systems analysis of the genetic interaction network of yeast molecular chaperones.

Molecular chaperones are typically promiscuous interacting proteins that function globally in the cell to maintain protein homeostasis. Recently, we had carried out experiments that elucidated a comprehensive interaction network for the core 67 chaperones and 15 cochaperones in the budding yeast Saccharomyces cerevisiae [Rizzolo et al., Cell Rep., 2017, 20, 2735-2748]. Here, the genetic (i.e. epistatic) interaction network obtained for chaperones was further analyzed, revealing that the global topological parameters of the resulting network have a more central role in mediating interactions in comparison to the rest of the proteins in the cell. Most notably, we observed Hsp10, Hsp70 Ssz1 chaperone, and Hsp90 cochaperone Cdc37 to be the main drivers of the network architecture. Systematic analysis on the physicochemical properties for all chaperone interactors further revealed the presence of preferential domains and folds that are highly interactive with chaperones such as the WD40 repeat domain. Further analysis with established cellular complexes revealed the involvement of R2TP chaperone in quaternary structure formation. Our results thus provide a global overview of the chaperone network properties in yeast, expanding our understanding of their functional diversity and their role in protein homeostasis.

Computational Analysis of the Chaperone Interaction Networks.

We provide computational protocols to identify chaperone interacting proteins using a combination of both physical (protein-protein) and genetic (gene-gene or epistatic) interaction data derived from the published large-scale proteomic and genomic studies for the budding yeast Saccharomyces cerevisiae. Using these datasets, we discuss bioinformatic analyses that can be employed to build comprehensive high-fidelity chaperone interaction networks. Given that many proteins typically function as complexes in the cell, we highlight various step-wise approaches for combining both the genetic and physical interaction datasets to decipher intra- and inter-connections for distinct chaperone- and non-chaperone-containing complexes in the network. Together, these informatics procedures will aid in identifying protein complexes with distinctive functional specializations in the cell that yield a very broad and diverse set of interactions. The described procedures can also be leveraged to datasets from other eukaryotes, including humans.