ALOHA: AI-guided tool for the quantification of venom-induced haemorrhage in mice

Venom-induced haemorrhage constitutes a severe pathology in snakebite envenomings, especially those inflicted by viperid species. In order to both explore venom compositions accurately, and evaluate the efficacy of viperid antivenoms for the neutralisation of haemorrhagic activity it is essential to have available a precise, quantitative tool for empirically determining venom-induced haemorrhage. Thus, we have built on our prior approach and developed a new AI-guided tool (ALOHA) for the quantification of venom-induced haemorrhage in mice. Using a smartphone, it takes less than a minute to take a photo, upload the image, and receive accurate information on the magnitude of a venom-induced haemorrhagic lesion in mice. This substantially decreases analysis time, reduces human error, and does not require expert haemorrhage analysis skills. Furthermore, its open access web-based graphical user interface makes it easy to use and implement in laboratories across the globe. Together, this will reduce the resources required to preclinically assess and control the quality of antivenoms, whilst also expediting the profiling of hemorrhagic activity in venoms for the wider toxinology community.


21
Abstract 22 Venom-induced haemorrhage constitutes a severe pathology in snakebite envenomings, 23 especially those inflicted by viperid species. In order to both explore venom compositions 24 accurately, and evaluate the efficacy of viperid antivenoms for the neutralisation of 25 haemorrhagic activity it is essential to have available a precise, quantitative tool for empirically 26 determining venom-induced haemorrhage. Thus, we have built on our prior approach and 27 developed a new AI-guided tool (ALOHA) for the quantification of venom-induced 28 haemorrhage in mice. Using a smartphone, it takes less than a minute to take a photo, upload 29 the image, and receive accurate information on the magnitude of a venom-induced 30 haemorrhagic lesion in mice. This substantially decreases analysis time, reduces human error, 31 and does not require expert haemorrhage analysis skills. Furthermore, its open access web-32 based graphical user interface makes it easy to use and implement in laboratories across the 33 globe. Together, this will reduce the resources required to preclinically assess and control the 34 quality of antivenoms, whilst also expediting the profiling of hemorrhagic activity in venoms for 35 the wider toxinology community. 36 37

38
Snakebite envenoming is a major public health problem, especially in the developing world 39 [1]. Indeed, it is responsible for substantial morbidity and mortality, particularly in the 40 impoverished areas of sub-Saharan Africa, South to Southeast Asia, Papua New Guinea, and 41 Latin America [1][2][3][4]. Whilst accurate estimates are difficult to make, it is believed that between 42 1.8-2.7 million people worldwide are envenomed each year, resulting in 80,000 to 140,000 43 deaths and 400,000 survivors left with permanent sequelae [5][6][7]. 44 45 The severity of a given envenoming is determined by several factors, such as the amount of 46 venom injected, the anatomical location of the bite, and the physiological status of the victim 47 [8]. In addition, there is a great variability in the composition of the venoms and the 48 predominant toxins present in different venoms, not only between genera, but also within a 49 single species [9]. Consequently, the clinical manifestations and pathophysiological effects of 50 envenomings can vary greatly depending on the offending snake species To allow for a standardised analysis of the haemorrhagic lesions, as well as to facilitate the 122 image analysis algorithms, we prepared an A4 printout sheet which the mice were placed on. 123 This sheet outlines where to place the mice and includes different lines and boxes of defined 124 lengths that allow for the scaling of the image (Fig.1). We also used a cut out mask to be 125 placed on the mice to facilitate lesion identification ( Fig.1). Printable versions of these two 126 components can be found in supplementary Figure 1 and 2. 127 128 129 Figure 1. Explanation of the printable template upon which the mice should be placed.

130
The multiple scales across the page allow for the automatic scaling by the tool and take into 131 account pictures obtained from different angles. The black frame acts as a cut out mask to 132 facilitate the automatic identification of haemorrhagic lesions. 133 134 2.4. Description of machine learning guided approach of quantifying 135 haemorrhagic activity 136 We trained a machine learning algorithm to automatically identify haemorrhagic lesions, adjust 137 for lighting biases, scale the image, extract haemorrhagic lesion area and intensity, and 138 calculate the HaUs. This was then implemented in an accessible fashion via a graphical user 139 interface (GUI) as the ALOHA tool (Fig. 2). 140 141 142 Figure 2. Overview of the workflow for ALOHA. First, the raw image is imported and 143 converted from sRGB to linear RGB. Thereafter, the image is white balanced and 144 subsequently further white balanced using the colour of the paper detected via the scaling 145 squares. In parallel, the image is rescaled using the same squares. This processed image is 146 then used for segmentation and automatic identification of the haemorrhagic lesions. 147 Together, this information is used to compute the lesion area and luminance, which is then 148 combined into a HaU score to assess the overall hemorrhagic lesion. 149 150 all images to ensure a resolution of 5 pixels per mm, which also allows for rapid computation. 212 213

234
In this study, we primarily present the results for the six images that used the template to 235 evaluate the performance of our fully automated method, ALOHA. We used, as a model, the 236 haemorrhagic lesions induced by the venom of B. asper on mice. 237 238

239
To address the potential impact of lighting differences, the tool automatically performs white 240 balancing. The white balancing works as expected and produces comparable results across 241 images (Fig. 3).

247
Using the scaling method outlined in 2.4.3, the tool automatically detects the image scales 248 and reports them (

255
To automatically identify lesion areas, the tool uses a machine learning guided segmentation 256 approach. Overall, an average MCC score of 0.8612 and an average F1 (Dice) score of 0.9064 257 was achieved, and we were able to predict 99.84% of the pixels correctly across 25 runs (Fig.  258  4).

268
To assess the severity of each lesion, the tool automatically computes the real-world area, 269 luminance, and HaU for each mouse in all of the test images (Tables 2,3

295
To ensure accessibility and easy implementation of ALOHA across research, production, and 296 quality control laboratories, a graphical user interface was developed ( 297 https://github.com/laprade117/ALOHA). Our tool can be used to quickly upload an image and 298 receive statistics on the lesion area, luminance, and HaU for each mouse in the image (

308
Haemorrhage is one of the key pathophysiological manifestations of snakebite envenomings, 309 particularly those inflicted by species of the family Viperidae [1,11]. Therefore, the preclinical 310 assessment of antivenom efficacy includes the evaluation of the neutralisation of 311 haemorrhagic activity [27]. This results in both a need for robust and reliable, but also rapid 312 assay approaches, while limiting required resources, which would allow their implementation 313 in diverse laboratory settings. In our earlier study, we were able to improve the classical rodent 314 skin test by increasing the accuracy of haemorrhage characterisation and reducing the time 315 required for analysis in comparison to the original approach, as well establish that HaUs 316 accurately reflect haemoglobin content analysis 18 . Therefore, our previous tool has been used 317 in a series of later publications [28][29][30][31][32][33][34][35][36][37]. Nevertheless, our prior approach still remained time 318 consuming, and required familiarisation with a new software and access to a colour pantone. 319 Furthermore, it was subject to human error as lesions were manually annotated, which 320 required training and would still differ from one individual to another. 321 322 In this study, we aimed to address these shortcomings by implementing a fully automated 323 analysis pipeline, aided by vision AI, i.e. U-Net. We found that our tool, ALOHA, was able to 324 rapidly and robustly assess the training images that covered a range of haemorrhagic lesion 325 severities. To ensure optimal accessibility and ease of implementation into existing workflows, we 336 developed a GUI-based web tool that allows users to conduct fast and accurate analyses of 337 haemorrhagic lesions. Using a smartphone, it takes less than a minute to take a photo, upload 338 the image, and receive accurate information on the severity of a venom-induced haemorrhagic 339 lesion in mice. This substantially decreases the analysis time required, from hours, to just a 340 few minutes. Furthermore, the ease-of-use significantly boosts the accessibility of our method 341 and provides a standard tool to be used across labs that does not require training or prior 342 knowledge on lesion assessment. 343 344 Despite the benefits ALOHA holds, some possible limitations may exist. In properly illuminated 345 environments, the white balancing performs accurately. However, in shadowed, or strangely 346 lit environments, the white balancing may not perform as well. To mitigate this, it is 347 recommended to photograph in bright, uniformly lit environments. It is especially important to 348 avoid casting shadows on or covering the black squares on the template paper for optimal 349 results. Furthermore, due to the translucency of a single sheet of paper, it is best to avoid 350 placing the template on a brightly colored table or desk when photographing. Scaling is 351 computed via information obtained from the black squares at the corners of the template 352 paper. This computation assumes that the black squares are perfect 10 mm x 10 mm squares. 353 Thus, for the most accurate results, it is best to use flat unwrinkled paper and photograph 354 directly from above, so that there is the least amount of distortion applied to the black squares.

377
The authors declare no conflict of interest.