The dataset consists\nof the video records of the rainbow trout in three cultivation units infected\nby Red mark syndrome disease (RMS). The videos were recorded by the digital\ncamera placed in the cultivation tank's waterproof box. Each video is 30\nminutes long. The records were taken during 3 months. Two tanks, TANK_01 and TANK_03\ncontain infected fish. TANK_20 contains control non-infected fish. The dataset\ndocuments the progress of the RMS visible signs on the fish skin from early to\nlate stages. The dataset was used to develop a remote early warning system for\nautomatic disease detection in aquaculture production. It can be used\nfor the studies in the filed of automation of aquaculture: fish individual identification,\ndisease analysis, fish behavior analysis, or fish detection.
Experimental animals
The rainbow trout (Oncorhynchus mykiss) used in the study\nwere hatched at DTU and raised under pathogen-free conditions. At a size of\naround 50 g, they were individually tagged with a passive integrated\ntransponder (PIT) tag in the dorsal musculature under anaesthesia. Before\nexperimentation, they were moved to the infection unit at DTU. The infection\nwas caused by cohabitation with RMS-affected seeder rainbow trout. The size of\nthe 29 seeders used for the experiment was 203\u00b126 g, 25\u00b11.2 cm. A total of\n180 healthy rainbow trout were used (160\u00b138 g, 22.9\u00b11.7 cm at the\nstart of the experiment and 348\u00b185 g, 29\u00b11.8 cm at the end on day 83). One\nhundred twenty were infected by cohabitation, and 60 were non-infected\ncontrols. The seeders were produced 61 days before the cohabitation with the\nexperimental fish. The cohabitation of the 29 seeders and 120 cohabitants\nwas from 20.9.2023 to 4.10.2023 when the seeders were\nremoved. The 120 cohabitants were subsequently divided into two\nexperimental tanks (tank TANK_01 and TANK_03). A third similar tank was\nused for 60 uninfected fish as a control group (tank TANK_20). See the detailed\nschedule in Figure 1. [PC1] All fish were kept in flow-through\n1m3 tanks supplied with a flow of fresh water at a rate of ~130 L/h. Water\ntemperature was 12.0+/-0.5 \u00b0C, the light regime was 12:12 h light: dark, and\nthe fish were fed commercial 3mm pellets at a rate of 1% total body weight\ndaily using clock feeders. One fish per tank was humanely euthanised at\ntwo-week intervals, and skin lesions were sampled for histological confirmation\nof RMS. The experiment was terminated on 12.12.2023. The experiment was\nconducted under the animal experimentation license 2019-15-0201-00159-C2 issued\nby the Danish Animal Experiments Inspectorate. The fish was sacrificed at the\nend of the experiment.
Figure 1. Experimental schedule.
RMS symptoms
RMS is a skin disease, and the primary clinical\nmanifestation is characteristic changes in the mucosal surface of the fish. The\ninitial symptoms are often subtle and may appear as lighter patches of skin.\nThese changes usually have a dorsoventrally elongated shape if found on the\nflanks. They gradually develop a pinkish hue and turn bright red in later\nstages (see Figure 2). Lesions are typically from 2 to 10 cm in diameter. Apart\nfrom the colour change (due to hyperemia/haemorrhaging), hallmark pathological\nsigns of RMS are oedema (raised skin) and scale loss (due to resorption) at\nmore advanced stages of lesion development. Ulcers can develop in severe cases,\noften as a result of secondary bacterial infections. This complicates the\ncourse of the disease and potentially results in fish mortality if therapeutic\nmeasures are not initiated in time (Galeotti et al., 2021), but it also\njeopardises a correct diagnosis. However, in the semi-controlled model at DTU,\nulceration is rare.
Figure 2. Examples of RMS visual symptoms. Left \u2013 early\nsymptom (6 weeks after cohabitation [PC2] ), light spot on the body,\nMiddle \u2013 the pinkish spot (8 weeks after cohabitation), Right \u2013 late symptom,\ndark red large lesion with missing scales (10 weeks after cohabitation).
Fish video records
The video records simulated the real fish cultivation\nconditions similar to those in fish production. Videos were recorded at regular\nintervals (every few days, 2-6 times per week) using an RGB digital camera\n(See3CAM_CU27) submerged in the housing in the fish tank, ensuring consistent\nmonitoring of symptom progression throughout the experiment, see Figure 1. The\nHD resolution of 1280 \u00d7 720 pixels and a frame rate of 80 fps were used\nfor video recording. The camera in housing was placed on a plastic tube holder;\nsee Figure 3. The camera was placed in a waterproof transparent plastic\nbox connected by an optical USB cable to the recording computer. The secure\nholder enables removing the camera from the tank during tank maintenance. The\nholder was equipped with a water-resistant LED light (20W, 5500K). Lighting was\nused to improve visibility in the tank, maximising overall image quality and\nincreasing camera framerate. Both the light and the camera were remotely\ncontrolled from the computer. The light was switched on only during the\nrecording period.
Figure 3. Left \u2013 The camera system consists of a plastic\nholder, an LED light, and the camera enclosed in a waterproof transparent box.\nRight \u2013 the top view of the tank with the fish.
All three tanks (two with infected fish and one control\ntank) were equipped with the camera system. The video recording started on\nOctober 9, five days after the seeders were removed. The earlier period was not\nrecorded because no RMS symptoms were expected (based on the knowledge of RMS\ndevelopment). The thirty-minute video records were taken during the whole\nexperiment for all tanks. Four video records were taken for the control tank,\nand forty video records were taken for two tanks with infected fish. The last\nvideo records were captured on December 10, two days before the last evaluation\nand experiment termination. In total, 44 video records (22 hours) of fish\nswimming were recorded. The fish freely moved in front of the camera. The\ncameras were focused at a distance greater than 15 cm to optimise the field of\nview and maintain focus on the majority of the fish swimming in the\ntank, Figure 4.
Figure 4. Examples of underwater video records from the\ncamera system. The two images demonstrate the difference in the illumination\nconditions.
Data preprocessing and annotation
Video recordings were preprocessed and manually annotated to\nprepare data for training the CNN-based detector. Providing ground-truth data\nenabled reliable localisation and identification of the progress of visual\nsigns on the fish skin at different stages. This process generated Video Image\nDataset (VID).
The VID images dataset was extracted from videos to create a\ndataset for model training from video records. All video records were manually\nchecked in the VirtualDub software, and the frame with visible signs on the\nfish was determined. When the sign was visible, the short track was used to\nextract every 10th frame. These frames represent the fish with visible\nsigns at different positions, orientations, and rotations and with light\nreflections on the fish\u2019s body. These parameters change during the fish\u2019s swimming,\nand the training dataset should consist of the images under the different\nconditions to generalise well. The same five classes used for still image\nannotation were used for video image annotation. The total number of annotated\nimages was 1991. The fish bodies were also annotated for the VID dataset to\nenable a more detailed analysis of the video record. The whole fish body\nrectangle, fish eye rectangle and dorsal fin rectangle were annotated for the\nfish with annotated signs of early or late RMS, see Figure 5. This triplet\nwas annotated only if all three parts were visible. In total, 1500 fish were\nannotated. The number of early RMS was 1426, and the number of late RMS was\n688.
Figure 5. Annotated images from short video sequence of\nearly RMS symptoms on one fish. The early RMS (green), fish (pink), eye (red),\ndorsal fin (orange) and snout/lower jaw (dark green) was annotated on all\nimages.
The\ndataset is stored in four ZIP files.
TANK_01.zip
TANK_03.zip
TANK_20.zip
Images from\nvideos.zip
Images and annotations.zip
\n\n
\n\n
TANK_01.zip -\ncontains 17 video (mp4 files). The name of the file represents the date and\ntime of the recording. Each file is 30 minutes long with the resolution of\n1270*720, 80fps, H264 codec.
\n - The file 20231129-190001.mp4 was recorded by different settings\n- 1920*1080, 60fps
\n - The file 20231130-220508.mp4 was recorded by\ndifferent settings - 1920*1080, 60fps
\n - The file 20231205-185541.mp4 was recorded by\ndifferent settings - 55 minutes long
\n - The file 20231206-194522.mp4 need not to be read properly
\n
\n
\n\n
TANK_03.zip - contains\n23 video (mp4 files). The name of the file represents the date and time of the\nrecording. Each file is 30 minutes long with the resolution of 1270*720, 80fps,\nH264 codec.
\n - The file 20231127-161610.mp4 was recorded by different settings\n- 1920*1080, 60fps
\n - The file 20231204-192209.mp4 was recorded by\ndifferent settings - 47 minutes long
\n - The file 20231107-174838.mp4 was recorded by\ndifferent settings - 46 minutes long
\n\n
\n\n
TANK_20.zip - contains\n4 video (mp4 files). The name of the file represents the date and time of the\nrecording. Each file is 30 minutes long with the resolution of 1270*720, 80fps,\nH264 codec.
\n - The file 20231024-171335.mp4 was recorded by\ndifferent settings - 59 minutes long
\n\n
Images from\nvideos.zip \u2013 contains three subdirectories: TANK_01, TANK_03, andTANK_20. Each\nsubdirectory contains subdirectories with the images selected from the particular\nvideo file of the tank. The name of subdirectory corresponds to the name of\nvideo file the images were selected from. The sub sub directory contain the\nimages in png image format. The description of image selection is provided in\nMetodologie part of the dataset description.
\n\n
Images and annotations.zip\n\u2013 it contains two folders:
- \nRMS only
- \nRMS+fish
\n\n
- \nEach folder contains images subfolder, labels subfolder and\nclasses.txt file. These files contain the all images selected form the video\nfiles and manually annotated by 6 or 9 classes. The format of annotation is YOLO\nbounding boxes. The description of image annotation is provided in Metodologie\npart of the dataset description.