The CT Sequence Selection Algorithm is most likely a specialised technique developed to increase the accuracy and efficiency of medical image processing, particularly for jobs such as tumor categorisation. It would be used to choose the most relevant CT scan slices or sequences from wider images. This is especially relevant in medical imaging because not all slices provide equivalent information for diagnosis or categorisation.We used a unique CT sequence algorithm before training the ResNet50V2-FPN model to reduce misclassification. The abovementioned technique yields results when the dataset is given as full sequence slices as input. The radiologist and physicians, on the other hand, can recognize the ovarian tumor anywhere the iliac crest bone appears in an imaging sequence. The iliac crest bone can be seen in the lower middle sequence of one CT scan image series. The model performs better and lowers false detection when trained with the lower middle image sequence. The CT sequence selection procedure must be used to delete picture slices from a sequence before using TIFF images to train a model.The CT Sequence Selection Algorithm was developed for several reasons. First, it was based on consultations with medical professionals who provided insights into the most diagnostically relevant CT sequences for ovarian tumors. This ensured that the algorithm focused on clinically meaningful sequences. Second, the software prioritised sequences that best indicate ovarian tumor morphology and contrast enhancement. The system selected high-quality photos without artefacts or low resolution to improve classification accuracy. The method ensured that selected sequences matched the tumor.The algorithm’s selection criteria were adjusted after several dataset tests and validation to improve classification precision. The tool’s scalable architecture, which works with several imaging modalities and disease types, also improves ovarian tumor categorisation. Finally, the method was tested and validated on several datasets to fine-tune its selection criteria, guaranteeing that it effectively improves classification precision. Because of its scalable architecture, the end product is a strong tool that increases ovarian tumour classification accuracy and can be adapted for other imaging modalities or cancer types.The flow diagram of the CT sequence selection algorithm is shown in Fig. 9.Figure 9Workflow diagram of CT Sequence Selection Algorithm.Analysis regionThe CT Sequence Selection Algorithm operates on a preprocessed TIFF image sequence as its input. Initially, the algorithm identifies a bounding box encompassing the iliac peak bone, referred to as the “Analysis Region.” This region is a reference for obtaining image slices containing the iliac peak bone within a sequence. To achieve this, the bounding box is positioned at the lower center of each image slice, ensuring pixel separation. This region must appear in the middle of every image slice consistently. However, the dataset comprises images from diverse cancer registry hospitals, each employing distinct measurement settings.Consequently, the images do not share a uniform scale, and the position of the peak bone varies among patients29. Thus, defining the Analysis Region necessitates meticulously assessing each sequence slice. Following several experiments, we established the Analysis Region dimensions as 220 to 370 pixels along the x-plane and 110 to 390 pixels along the y-plane ([220,370] to [110,390]). These dimensions effectively range from larger to smaller iliac crest bones across slices. Within this region, pixel data corresponding to the lower center area of the ovarian tumour in each image slice is contained. An exemplar image slice, complete with a loaded alpha channel and the designated Analysis Region, is illustrated in Fig. 10 (a-c).Figure 10a Image slice with Analysis Region. Original Image b) Alpha channeled Image c) Analysis Region.Every image includes a fourth channel, which is called the alpha channel. It describes the transparency of the picture. It can be obtained by passing the alpha value to an original image. The alpha value selected here is -1.Image segmentationThe HSV color space34 characterizes color through three components: Hue (H), Saturation (S), and Value (V). Represented as a hexagon, Hue is denoted by an angle within the range of [0, 2𝜋]. Saturation quantifies color’s purity or depth, while the value assigns a specific grey shade to which transformations converge35. In contrast, the RGB color space involves components R (Red), G (Green), and B (Blue), which correspond to color intensity. RGB images intertwine color information with color intensity, making their separation unattainable. HSV overcomes this limitation by isolating color information from image intensity, rendering it well-suited for segmentation and object-tracking purposes. Each slice comprises RGB color channels within a sequence, with the object of interest being the iliac crest bone. A segmentation process facilitates object tracking, converting RGB representation into the HSV color space36.Steps to follow in the conversion of RGB to HSV:
Step 1: RGB values are divided by 255 to change the range from 0…0.255 to 0. 1 defined as:
R′=𝑅⁄255
G′=𝐺⁄255
B′=𝐵⁄255
Step 2: Calculate the maximum and minimum channel value defined as:
Cmax=max (R′, G′, B′)
Cmax=max (R′, G′, B′)
Δ=Cmax−Cmin
Step 3: Compute Hue, Saturation, and Value using the formula below: Hue calculation:$$\:H=\left\{\begin{array}{c}{0}^{0}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\varDelta\:=0\\\:{60}^{0\:}\times\:\left(\frac{{G}^{{\prime\:}\:}-{B}^{{\prime\:}}}{\varDelta\:}mod\:6\right),\:\:\:{C}_{max}={R}^{{\prime\:}}\\\:{60}^{0\:}\times\:\left(\frac{{B}^{{\prime\:}\:}-{R}^{{\prime\:}}}{\varDelta\:}+2\right),\:\:\:\:\:\:\:\:{C}_{max}={G}^{{\prime\:}}\\\:{60}^{0\:}\times\:\left(\frac{{R}^{{\prime\:}\:}-{G}^{{\prime\:}}}{\varDelta\:}+4\right),\:\:\:\:\:\:{C}_{max}={B}^{{\prime\:}}\end{array}\right.$$
Saturation calculation:
$$\:S=\left\{\begin{array}{c}\:\:\:\:\:\:0\:\:\:\:\:\:\:\:\:\:\:,{C}_{max}=0\:\:\:\\\:\frac{\varDelta\:}{{C}_{max}}\:\:\:,{C}_{max}\ne\:0\end{array}\right.$$
Value calculation:
V=Cmax
Figure 11a and b show image segmentation and enhancement using HSV color space for a crest bone slice in an Analysis Region.Figure 11Segmentation using HSV color space. a) Original Image b) HSV color Space image.In the next step, Select the range (0,0,0) as dark and (250,250,250) as light and apply the HSV color space image (Fig. 10b) to find the masked image, which is the iliac crest bone part. The BitwiseAnd(&) operation is applied to the mask and original image slice (Fig. 10a) to get the required result, as shown in Fig. 12a and b.Figure 12Result after BitwiseAnd operation. a) Masked Image b) BitwiseAnd image.Crop the analysis region from a masked image with region [220:370,110:390] as shown in Fig. 13.Figure 13Analysis Region of the masked image.Applying the mask to the input image slice transforms dark pixels into white and vice versa. This phenomenon is depicted in Fig. 13, where, within an analysis region, the boundary of the iliac crest bone is delineated by dark pixels, while all other pixels are converted to white. This characteristic implies that the presence of the iliac crest bone in a slice is determined by the count of dark pixels within the analysis region. The images within our dataset are captured as 16-bit grayscale images. In a 16-bit grayscale image, pixel values range from 0 to 5000. Specifically, pixel values between 0 and 200 are categorized as dark pixels. Subsequently, upon selecting the analysis region, the subsequent step involves enumerating the dark pixels with intensity values below 200 within this region. This pixel-counting process applies to every image slice in a sequence.For instance, consider the analysis mentioned above the region ([220,370] to [110,390]), which spans 150 × 280 pixels, totaling 40,000 pixels. Within this, 5300 pixels possess values below 200, while the remaining 36,700 pixels surpass the 200-pixel threshold. The procedure involves presenting a sequence of slices as individual inputs to the CT sequence selection algorithm. Each slice’s analysis region is designated, and the count of dark pixels below the 200-pixel threshold is calculated within each region. This iterative process is applied to every slice within a sequence. To identify slices containing the iliac crest bone in a sequence, it becomes imperative to establish a threshold, a topic elaborated upon in the subsequent subsection.Threshold setupTo fix a threshold, initially scan all the slices in a sequence and count the dark pixels in an analysis region12. Suppose there are 160 slices in one sequence, and the number of dark pixels of each slice is shown below in an array.[2916, 2447, 3695, 3896, 4064, 3738, 3814, 4047, 4138, 4664, 5028, 5249, 2194, 5300, 5317, 5466, 5580, 5509, 5270, 5139, 5020, 5050, 5183, 2586, 5066, 4948, 4846, 4587, 4276, 3903, 3423, 3372, 3282, 3211, 1989, 3231, 3409, 3741, 4017, 4147, 4103, 4429, 4614, 4693, 5093, 2417, 5381, 5694, 5956, 6037, 6169, 6094, 5962, 5490, 5090, 4603, 2241, 4574, 4176, 3914, 3947, 4120, 4080, 3930, 3741, 3374, 3013, 2223, 2683, 2538, 2456, 2130, 2831, 2452, 2022, 2588, 2081, 2174, 2110, 1807, 1739, 2297, 2181, 2492, 2681, 2559, 2365, 2227, 1969, 1842, 2043, 1591, 1714, 1825, 2879, 1637, 1814, 2000, 2054, 1970, 1841, 1803, 1587, 1644, 1544, 2700, 1522, 1538, 1730, 1825, 1667, 1622, 1328, 1810, 1668, 1693, 2670, 1685, 1558, 1393, 1484, 1300, 990, 1120, 1556, 1551, 1444, 2786, 1637, 1358, 1115, 880, 1407, 1087, 763, 1198, 1775, 1450, 2483, 1393, 1343, 1496, 1227, 974, 1312, 1867, 917, 907, 1687, 2550, 1701, 1444, 1694, 1943, 1899, 2098, 2474, 2433, 3247, 3583]In this array, maximum and minimum dark pixel counts are considered for computing a threshold, which is defined in Eq. 6$$\:\text{T}\text{h}\text{r}\text{e}\text{s}\text{h}\text{o}\text{l}\text{d}=(\text{M}\text{a}\text{x}\_\text{d}\text{a}\text{r}\text{k}\text{P}\text{i}\text{x}\text{e}\text{l}\text{C}\text{o}\text{u}\text{n}\text{t}-\text{M}\text{i}\text{n}\_\text{d}\text{a}\text{r}\text{k}\text{p}\text{i}\text{x}\text{e}\text{l}\text{c}\text{o}\text{u}\text{n}\text{t})/2.3$$
(6)
The threshold is fixed by the trial-and-error method; instead of taking the average of two, it is fixed to 2.3 to select more image slices and to choose a small iliac crest bone image slice.In the above example, the dark pixel count array of 160 slices Max_darkPixelCount is 6169, and Min_darkPixelCount is 763. The threshold becomes:Threshold = (6169 − 763)/2.3 = 2350.4347.The imaging scale differs from one sequence to another. Fixing one threshold for the entire dataset containing a set of sequences is impossible. Hence, the threshold is calculated for each sequence.Selection of iliac crest bone slicesIn this step, selecting a slice that contains an iliac crest bone and discarding the remaining slices is carried out. For this purpose, the calculated threshold is used. If the dark pixels of the Analysis Region are greater than a threshold, select such slice containing iliac crest bone. If the dark pixels of the Analysis Region are less than a threshold, discard such a slice that does not have iliac crest bone. For the above example of 160 slices, the selected slices’ dark pixel array, which is greater than the threshold, is:[2916 2447 3695 3896 4064 3738 3814 4047 4138 4664 5028 5249 5300 5317.5466 5580 5509 5270 5139 5020 5050 5183 2586 5066 4948 4846 4587 4276.3903 3423 3372 3282 3211 3231 3409 3741 4017 4147 4103 4429 4614 4693.5093 2417 5381 5694 5956 6037 6169 6094 5962 5490 5090 4603 4574 4176.3914 3947 4120 4080 3930 3741 3374 3013 2683 2538 2456 2831 2452 2588.2492 2681 2559 2365 2879 2700 2670 2786 2483 2550 2474 2433 3247 3583]Among 160 slices, the total number of image slice selected are 84 slices. The remaining image slices are discarded. Each sequence is captured from the abdomen to the pelvis. By applying the CT Sequence Selection Algorithm, 50% of image slices that do not contain ovarian tumours are discarded. This procedure repeats for each sequence. The complete workflow of CT Sequence Selection Algorithm is written in the form of algorithm 1.1, as shown below:Algorithm 1.1. CT Sequence selection algorithm.