An Online AI-Powered Interactive Histological Image Annotation Platform for Analyzing Intestinal Regenerating Crypts in Post-Irradiated Mice

PURPOSE/OBJECTIVE(S)

The goal of this project is to build an online AI-powered interactive annotation platform to accurately and efficiently annotate intestinal regenerating crypts in histological images of mice after abdominal irradiation.

MATERIALS/METHODS

The proposed platform is developed by the seamless integration of a front-end web client and a back-end server. Such client/server design allows the users to access the platform without software installation on local computers. Our front-end client is developed with SvelteJS + WebGL technology stack, allowing access from any common web browsers and enabling user interaction, such as image importing/visualization, interactive crypt annotating, and annotation saving/deleting. The back-end server is responsible for executing the tasks requested from the web client, for instance, image pre-processing, AI-based crypts automatic identification, and database management. The image preprocessing is designed to extract a single cross section image using morphological operations because multiple hematoxylin and eosin (H&E) stained jejunum cross sections from post-irradiated mice are scanned within one slide. The auto-crypt identification is powered by a trained and validated AI engine U-Net, classifying image grid tiles into two groups with and without regenerating crypts. The database is implemented with the self-contained SQLite to support recording and indexing the annotated grid tiles with regenerating crypts. The workflow for crypt analysis on this interactive platform has 5 steps: 1) manually import a whole H&E slide image; 2) auto-preprocess the slide by extracting single cross-section images; 3) auto-identify regenerating crypts with an AI engine; 4) interactively annotate (add, delete, modify) auto-identified crypt markers; 5) save and/or output the annotation to the database or the local drive.

RESULTS

The performance of the developed interactive crypt analysis platform was evaluated in aspects of accuracy and efficiency. The AI-powered crypt auto-identification accuracy was assessed by computing the mean absolute error (MAE) on crypt number per cross section between manual and auto annotation using a testing dataset containing 80 cross sections. It achieved an MAE of 3.5±4.8 crypts per cross section, and 81.25% of the cross sections have no more than 5 crypts difference. The efficiency was assessed under two conditions with the server on the cloud and a local computer. It took about 2-3 minutes to finish the entire workflow on the cloud, while 1-2 minutes on the local by saving ∼1 minute on image uploading.

CONCLUSION

The developed web client/server platform enables online automatic identification and interactive annotation of mice crypts in minutes. It is a convenient tool that allows accurate and efficient crypt analysis and can be extended for other histologic image analyses.

Deep Learning-Based Pipeline for Automatic Identification of Intestinal Regenerating Crypts in Mouse Histological Images

PURPOSE/OBJECTIVE(S)

A classical approach for evaluating normal tissue radiation response is to count the number of intestinal regenerating crypts in mouse histological images acquired after abdominal radiation. However, manual counting is time-consuming and subject to inter-observer variations. The goal of this study is to build a deep learning-based pipeline for automatically identifying intestinal regenerating crypts to facilitate high-throughput studies.

MATERIALS/METHODS

Sixty-six healthy C57BL/6 female mice underwent 16 MeV whole abdominal electron irradiation. The small bowel was collected from each mouse 4 days post-irradiation, and 9 jejunal cross-sections from each were processed together in a single slide. The slides were stained with hematoxylin and eosin (H&E) and subsequently scanned (x20), providing one electronic histological image per mouse. Regenerating crypts, consisting of more than 10 basophilic crypt epithelial cells, were manually identified using point annotations in histological images. The pipeline was built to take the input of the image containing 9 cross sections and automatically identify the regenerating crypts on each cross section. It mainly consists of two components, cross section segmentation using intensity thresholding and morphological operations and crypt identification using a UNet. The dataset was randomly split into 46, 10, and 10 slide images for UNet training, validation, and testing. Each slide image was split into grid tiles with a voxel size of 200 × 200, and 40 × 40 square masks were placed with centers at manual point annotations on tiles with regenerating crypts. 5203/5198 tiles (w/wo crypt mask) were extracted to train UNet by minimizing dice loss. The mask probability map generated by the UNet was post-processed to identify the crypt position. Postprocessing hyperparameters were tuned using the validation dataset. The model accuracy was evaluated using the testing dataset by computing the mean absolute error (MAE) of the crypt number averaged across all cross sections.

RESULTS

The number of regenerating crypts on testing cross sections ranges from 1 to 63. The testing cross-section-wise MAE achieved by the platform is 3.5±4.8 crypts. 81.25% of testing cross sections have absolute number differences less than or equal to 5 crypts.

CONCLUSION

Our established deep learning-based pipeline can accurately count the number of regenerating crypts in mouse intestinal histological images. We have integrated it into an online platform that enables automatic crypt identification and allows users to interactively modify auto-identified crypt annotations. The acquired annotations from the platform will be used to finetune the deep learning model to achieve better identification performance.

Equivalent Dose Estimation in FLASH Irradiation with a Deep Learning Approach

PURPOSE/OBJECTIVE(S)

Ultra-high dose rate (FLASH) irradiation has been reported to provide decreased normal tissue toxicity without compromising tumor control compared with conventional (CONV) irradiation. However, a comprehensive understanding of the FLASH biological effect requires precise quantification of radiobiology. The study is to explore whether deep learning (DL) can tackle the task. As a proof of concept, we investigate a DL model for estimating FLASH dose to its equivalent CONV dose.

MATERIALS/METHODS

Healthy C57Bl/6 female mice underwent FLASH (200Gy/s; n = 43) or CONV (0.12Gy/s; n = 41) whole abdominal irradiation using ∼16 MeV electron beams with a dose escalation scheme of 5 groups (n = 8 or 9) at 1Gy increments: 12-16Gy FLASH, 11-15Gy CONV. 4 days post-irradiation, 9 jejunum cross-sections per mouse were H&E stained for histological analysis. Each cross-section image was processed to remove lumen background and oversampled into multiple large-scale and small-scale patches along jejunal circumference. In CONV dataset, we randomly selected the data of 32 mice (80%) for model training and the rest (20%) for model validation. A ResNet101-based DL model, pre-trained with an unsupervised contrastive learning scheme, was retrained with only CONV training set to estimate corresponding CONV dose. For comparison, a crypt counting (CC) approach was implemented by manually counting the number of regenerating crypts on each cross-section image. An exponential function of dose vs crypt number was fitted with the CONV training set and used for dose estimation on the testing set. Mean squared error (MSE) was used to assess the accuracy of DL and CC approaches in estimating dose levels in CONV irradiation. The validated DL model was applied to the FLASH set to project FLASH dose into corresponding CONV dose that results in equivalent biological response.

RESULTS

The CONV dose estimated by DL and CC approaches and DL-estimated FLASH equivalent dose were summarized in Table 1. The DL model achieved an MSE of 0.21 Gy2 on CONV testing set compared with 0.32 Gy2 of the CC approach. FLASH equivalent dose estimated by DL model for 12, 13, 14, 15 and 16Gy were 12.16±0.40, 12.53±0.32, 12.72±0.24, 12.85±0.20 and 13.04±0.27 Sv, respectively.

CONCLUSION

Our proposed DL model can accurately estimate the CONV dose based on histological images. The DL predictions of FLASH dataset demonstrate that FLASH may reduce normal tissue toxicity with a lower equivalent dose, especially at high irradiated dose levels. Our study indicates that deep learning can be potentially used to assess the equivalent dose of FLASH irradiation to normal tissue.

Anatomically Realistic 3D Printed Mouse Phantom for Multi-Institutional Benchmarking of FLASH and CONV Irradiation

PURPOSE/OBJECTIVE(S)

It is reported that about US$28B/year is spent on pre-clinical studies that are not reproducible. FLASH studies may suffer from the same reproducibility crisis due to the non-standard nature of the FLASH beamlines and the lack of dosimeters that can function at ultra-high dose-rates. There have been reports of different outcomes with regard to the FLASH effect across different institutions, even though similar beamlines, temporal structure, and nominal dose levels were used. This brings up the question of the accuracy of dosimetry under FLASH conditions for a fair comparison between FLASH and CONV. To answer this question, we develop and characterize an anatomically realistic 3D-printed mouse phantom to be used in a multi-institutional dosimetric benchmarking effort.

MATERIALS/METHODS

Mesh files for bony anatomy, lungs, and soft tissue derived from a CT scan of a mouse were converted to an editable 3D model. The 3D model was cut along the coronal plane and modified to allow the inclusion of radiographic film. A multi-material approach was employed to print the phantom. A dual-nozzle 3D printer was used, where one of the nozzles used Acrylonitrile butadiene styrene (ABS) to mimic soft tissue and the other nozzle used Polyactic acid (PLA) to mimic bone density. The two materials were used together in a single print. Lungs were approximated by lightweight PLA and were printed separately and inserted into corresponding cavities in the phantom. Hounsfield Units (HU) and print-to-print stability were verified. Radiographic films were laser cut for different anatomical sites. Two institutes took part in this study with data pending from 3 more institutions. The institutes were instructed to deliver 10 Gy to the plane of the film for the whole abdomen, whole lung, and brain irradiations. 2D dose maps were compared between FLASH and CONV, and the deviation from the prescribed dose was also measured.

RESULTS

The 3D-printed soft tissue, bone, and lung densities were measured to be ∼ 1.01 g/cc, 1.22 g/cc, and 0.44 g/cc, respectively. For soft tissue and bone, the Hounsfield unit (HU) difference from one print to another was < 10 HU. The greatest variation was within the lungs (54 HU), but this had a minimal effect on the dose distribution (<1%). For the two institutions that completed the survey, the maximum average difference between FLASH and CONV for all irradiations was 0.75 Gy (7.48%). The maximum average difference from the prescribed dose for all irradiations was 0.7 Gy (7.20%) across both institutions. The largest discrepancy was generally observed to be for lung irradiation, indicating that lack of treatment planning systems limits our ability to prescribe accurately in areas of inhomogeneities.

CONCLUSION

A 3D printed anatomically realistic mouse phantom was developed, characterized, and used in a multi-institutional dosimetric benchmarking effort. Such a study is paramount for the clinical translation of FLASH as it facilitates reduced variability from one institution to another.

Comparison of Tumor Control between FLASH and CONV in an Orthotopic Breast Cancer Model

PURPOSE/OBJECTIVE(S)

Post-lumpectomy radiotherapy (RT) reduces in-breast tumor recurrence by eradicating residual, occult breast cancer (BC) that may be in the mm size scale. The ability of FLASH-RT to eradicate BC relative to conventional dose rate (CONV) RT is unknown. ∼ 20Gy RT is currently used clinically for single-fraction breast IORT. Determine the effectiveness of FLASH compared to CONV in eradicating small tumors in an orthotopic, syngeneic model of BC using single-fraction 20 or 30Gy RT.

MATERIALS/METHODS

Radiation sensitive, mammary tumor cell line Py117 from the transgenic model of the mouse mammary tumor virus promoter driving the polyoma middle T antigen (MMTV- PyMT) efficiently forms non-metastatic, orthotopic tumors in C57BL/6 mice. 106 Py117 cells were injected orthotopically into the left 4th mammary fat pad of C57Bl/6J mice. Radiotherapy was performed with a custom jig that allows for fixed positioning of the target volume (2x2cm radiation field) with 5mm of margin into surrounding tissue. Tumors were irradiated at ∼30mm3 volume or, for comparison, at a range of greater volumes (200-800mm3) with 20 or 30Gy FLASH or CONV with 16-17 MeV electrons.

RESULTS

Small 30mm3 tumors regressed until ∼ day 15 after 20Gy single fraction RT then regrew for both FLASH and CONV. 30mm3 tumors were eradicated with both FLASH and CONV at 30Gy with no regrowth up to day 35 post-RT. Larger tumors irradiated with 30Gy regressed until ∼ day 12 post-RT then regrew for both FLASH and CONV. There was no significant difference in growth delay or tumor eradication between FLASH and CONV in any cohort.

CONCLUSION

FLASH was as effective as CONV in controlling growth and eradicating murine BC. Based on other preclinical studies, single-fraction doses between 20 and 30Gy, as well as hypofractioned RT schedules, may identify FLASH doses that achieve comparable tumor control with less toxicity than CONV. Such findings would encourage clinical trials of FLASH in human BC.

Heavy Metal Bioaccumulation in the Anemone Paraphelliactis pabista Dunn, 1982 (Actiniaria: Hormathiidae) from the Hydrothermal System of Guaymas Basin, Gulf of California

A single specimen of the anemone Paraphelliactis pabista was recovered from the Southern Trough of Guaymas Basin during the deep-sea expedition Extreme 2008 conducted onboard the R/V Atlantis/DSRV-2 ALVIN. We studied the bioaccumulation capacity of heavy metals in various tissues of the anemone (oral disk-columella-pedal disk), and retention or adhesion of mineral particles in the epidermis, mesoglea, and gastrodermis. The digested tissues were analyzed for As, Ba, Co, Cu, Cr, Fe, Mn, Ni, Pb, Se, Sb, Sr, Ti, V, and Zn by inductively coupled plasma mass spectrometry. This analysis revealed the capacity of P. pabista for accumulating heavy metals. The predominant mineral particles identified in tissue samples was barite followed by Fe, aluminum-silicates, Sr, and with less presence Cr, Ti, and pyrite. Of the three body compartments analyzed of this anemone, the oral and pedal disks show a greater capacity of bioaccumulation of heavy metals than the columella.

Epibenthic amphipod abundance and predation efficiency of the pink shrimp Farfantepenaeus duorarum (Burkenroad, 1939) in habitats with different physical complexity in a tropical estuarine system

Amphipod abundance and biomass were determined in soft-bottom substrates (SBS), monospecific Thalassia testudinum patches and T. testudinum with attached macroalgae (SAV) from Términos Lagoon. Amphipods were absent in SBS, and their density and biomass were higher in SAV (3351 individualsm(-2), 1718 mg AFDWm(-2)) than in T. testudinum (1220 indm(-2), 625 mg AFDWm(-2)). Although macroalgae and seagrasses are recognised as an alternative refuge against predation for amphipods, the high abundance of amphipods in SAV suggests that macroalgae represent a habitat that provides greater food availability. Pink shrimp Farfantepenaeus duorarum (Burkenroad, 1939) consumption rate (Mo) of epibenthic amphipods was experimentally evaluated. Mo intensifies as prey density increases and varied from 0.39 to 2.39 mg AFDWh(-1). Predation efficiency of F. duorarum on epibenthic amphipods was also evaluated in four artificial habitats with different physical complexity: soft-bottom substrates (SBS), small woody debris (SWD), seagrasses with densities of 300 and 1200 shootsm(-2) (S300 and S1200, respectively), macroalgae (MA), and at two prey densities (962 and 2406 indm(-2)). Amphipod consumption rate by F. duorarum varied from 1.20 to 2.07 indh(-1) in S1200 and MA, respectively. Habitat complexity had a significant effect on consumption rate, but prey density did not. Habitat physical complexity and predation efficiency maintained an inverse and a non-linear relationship. Presumably, the decrease in predation efficiency in association with the habitat complexity is due to the differential refuge value of these habitats. However, predation efficiency may also be influenced by either the microhabitat use by amphipods, the shrimp's dependence on seagrasses, or by differences in habitat value caused by the diel behavioural distribution pattern of amphipods and shrimp. Both field and experimental results highlight the importance of evaluating the relative value of tropical estuarine habitats through the study of the relationship between habitat physical complexity and predator-prey interactions. They also emphasise that inherent biological and ethological factors of the predator and prey involved, coupled to spatial and temporal variations in the habitat, should also be considered.