Proof of Concept Clinical Evaluation of Gradient Different Weight Threshold Segmentation for Accurate Tumour Delineation in I-131 SPECT Imaging

INTRODUCTION

Accurate tumour segmentation is critical for patient-specific dosimetry in low-dose personalised ablative therapy (LDPA) using I-131.^[1-3] Precise delineation of tumour boundaries affects dose calculations and treatment planning, directly influencing therapeutic outcomes. Improved segmentation accuracy helps target the prescribed radiation dose to the tumour while minimising exposure to adjacent healthy tissues, thereby supporting treatment efficacy and patient safety.

In single-photon emission computed tomography (SPECT) with Iodine-131, accurate tumour segmentation remains challenging due to the modality’s inherently low spatial resolution^[4,5] and the resulting unclear tumour boundaries.^[6,7] These imaging limitations make it difficult to delineate lesions from surrounding tissue in SPECT images. Numerous segmentation techniques have been explored in this context.

Traditional threshold-based methods (for example, using a fixed percentage of the lesion’s maximum intensity) are simple and widely used,^[8] but their accuracy is inconsistent. The optimal threshold strongly depends on the tumour-to-background contrast, and a single fixed value often fails to generalise across different cases.^[9-11]

More advanced approaches, including machine learning and deep learning models, have also been applied to SPECT tumour delineation; however, such models typically require extensive training data and may perform unreliably on “unseen” data, limiting their generalisability in clinical practice.^[12] These shortcomings motivate the development of more robust, accurate segmentation methods for SPECT I-131 images.

In this work, we propose a tumour segmentation approach based on a modified gradient-different-weight threshold technique (implemented in MATLAB), adapted from the lower limit window level (LLWL) method described by Li et al.^[13] The approach uses gradient-derived voxel weights better to handle blurred edges and low-contrast boundaries in SPECT. In addition, the threshold selection is linked to a measurable local ratio factor (Cmax/Cmean) and calibrated using a NEMA IQ phantom, providing a structured alternative to operator-dependent fixed-percentage thresholding. We report phantom optimisation and a proof-of-concept clinical evaluation using two retrospective I-131 SPECT/CT datasets.

MATERIAL AND METHODS

As shown in Fig 1, the gradient-difference weighting strategy assigns a weight to each voxel by considering the local gradient magnitude, producing a voxel-weighting array. In the MATLAB implementation used here, grey-difference weights quantify intensity differences that help separate foreground (target) from background.^[14-16] The reference intensity for weighting was linked to the local ratio factor computed from the maximum voxel count within the target and the estimated mean background.

Close

Fig 1: Gradient different weight threshold method process flow. SPECT: Single-photon emission computed tomography

Export to PPT

The significance of each voxel correlates with the gradient magnitude at its location. Voxels with low gradient magnitudes (smooth regions) have higher weights, while voxels with high gradient magnitudes (edges) have lower weights.^[17] Gradient weighting uses sigma as the standard deviation of the Gaussian derivative, which is used to compute the image gradient (MathWorks Help Centre, 2021). In this study, the voxel-weighting values are used to derive a threshold value that is proportional to the gradient difference between the foreground and background.

Modification of the gradient different weight threshold method

In this study, two steps were taken to modify the gradient different weight threshold method based on the gradient different weight syntax in MATLAB 2023b software. First is the iteration of the threshold value for the known spherical volume in the NEMA IQ phantom. Second, the determination of the ratio factor is representative of the sphere-to-background ratio. Fig 1 shows the process of changing the gradient with a different weight threshold method.

Based on this syntax, the grey different cut-off value should be the voxel value at the boundary of the target volume to be segmented, as shown in Fig 2. However, it is challenging to determine the grey different cut-off value for the grayscale image of the SPECT image because the boundary difference from the background is similar, especially because SPECT provides low-resolution images.

Close

Fig 2: Voxel counts in the grayscale image (arrow).

Export to PPT

To counteract this problem, the syntax for the gradient with different weights was changed, i.e. the value 'GrayDifferenceCutoff ' was set to a constant value of 1,000,000. This value was chosen because the voxel number in the SPECT image can never be reached. Nevertheless, the threshold value for the gradient's different weight thresholds must be investigated and calibrated in order to obtain a highly accurate volume segmentation. In this study, the NEMA IQ phantom with six spheres is used to determine the correlation between threshold and ratio factor. The sphere to background ratios of 12:1, 10:1, 8:1 and 6:1 were included to obtain the ratio factor.

The formula below was used to obtain the ratio factor's value:

where Cmax refers to the highest number of pixels in the sphere, while Cmean corresponds to the mean number of pixels for the background. To obtain the best mean background values of the NEMA IQ phantom, three methods were used as described in the results.

Techniques of mean background tumour

Three methods were chosen to determine the mean background of the NEMA IQ phantom. The first technique is to draw 12 circles on the background area of the phantom, as shown in Fig 3. The diameter of each circle is a constant of 37 mm. Then, the sum of all total counts in each circle was divided by the number of circles drawn to obtain the mean background count.

Close

Fig 3: Circle mean background counts. STD: Standard deviation

Export to PPT

The second technique, a narrow annular background region, is adjacent to the volume, as shown in Fig 4. MATLAB software with the “imtool3D” function was used to draw circles and perform statistical counts to obtain the optimal average background.

Close

Fig 4: Narrow annular mean background. STD: Standard deviation

Export to PPT

The narrow annular background was drawn based on its visual characteristics, and numerous studies have used this approach. To determine an accurate ROI, an appropriate ULWL value was used. The ULWL value was determined using the following formulas:

Here, t denotes the total number of counts for the slice with the highest count, p stands for the number of pixels, and m remains a constant (set to 10 in this study), as shown in Fig 5. In this study, a 128 × 128 matrix was used for all slices, resulting in a total number of 16384. The correct choice of the parameter m is crucial to ensure the delineation of the target tissue on the colour display. For this study, the m value of 10 was chosen, although alternative values such as 8, 12, 14, 16 and 18 could also be reliable. Consistency in the selection of m is critical for overall data processing and ROI delineation. Different values of m were tested (from 8 to 18), and different results were obtained for the images.

Close

Fig 5: Tranverse slice of maximum voxel count image with different m value ; (a) m=8; (b) m=10; (c) m=12; (d) m=14; (e) m=16; (f) m=18

Export to PPT

For the third technique, the full wave with half maximum (FWHM) was used, where each sphere was drawn with four long lines through the central voxel maximum of each sphere, as shown in Fig 6. Each line was plotted in the bell curve profile, exported to the workspace of MATLAB, and then the FWHM was calculated from the MATLAB file exchange program as shown in Fig 7. The FWHM for each line is calculated to obtain a boundary for the narrow ring to be drawn, as shown in Fig 8.

Close

Fig 6: Four lines to get the bell curve for full wave with half maximum Calculation: a) Bell curve for first sphere; (b) Bell curve for second sphere; (c) Bell curve for third sphere; (d) Bell curve for fourth sphere

Export to PPT

Close

Fig 7: Full wave half maximum (FWHM) calculation.

Export to PPT

Close

Fig 8: Mean background counts using full wave have maximum methods; (a) Sphere 1; (b) Sphere 2; (c) Sphere 3; (d) Sphere 4 FWHM: Full wave with half maximum, STD: Standard deviation

Export to PPT

To obtain the optimal threshold value for each sphere, the threshold value is automatically determined based on the required volume with the looping script based on the gradient, different weight syntax and fixed volume of the sphere. Based on equation (1), the ratio factor is based on the division of the maximum count by the mean background count. The maximum counts used in this calculation are due to the fact that each target volume to be segmented has a maximum point. This is based on radon transformation theory, where each detected gamma represents a specific point in a projection angle, i.e. the ideal line integrals of the radioactive source distribution. This is the basis for why the maximum counts are used to calculate the ratio factor.

Graphical user interface gradient with different weights

A custom MATLAB (R2023b) graphical user interface (GUI), as shown in Fig 9, was developed to implement the modified gradient different weight threshold segmentation method. Grey difference weights were computed using GrayDiffweight with a fixed GrayDifferenceCutoff (1,000,000), and tumour masks were generated using image segmentation using fast marching method with the threshold selected from the GUI based on the ratio factor. The GUI provides synchronised axial/coronal/sagittal navigation, colour-labelled multi-tumour overlays, a region statistics table, and a 3D CT– tumour fusion view for visual validation and manual fine positioning of individual tumour surfaces.

Close

Fig 9: Graphical User Interface gradient with different weights

Export to PPT

Orthogonal 2D views (axial/coronal/sagittal) display the SPECT/PET volume with colour overlays and tumour labels, while the 3D fusion panel renders the resampled CT isosurface together with individual tumour surfaces. The control panel supports background/seed selection, threshold mode selection, segmentation execution, region-based quantification (regionprops3), and 3D interaction (rotate/zoom and per-tumour translation).

RESULTS

The optimisation of the gradient different weight threshold method is based on the three types of techniques for determining mean background counts for calculating the ratio factor. Meanwhile, the threshold value is the same for all four background ratios from sphere one to sphere five. Although the NEMA IQ phantom has six spheres, only the ratio factor and threshold value of sphere one to sphere five are under-considered since the total counts in sphere six are almost the same as the mean counts in the background. Thus,sphere six in this study was negligible. All the ratio factors and threshold values are shown in Table 1.

Fig 10-12 show the graph ratio factor versus threshold value plotted for the three types of mean background techniques. Among the three graphs plotted, it can be noted that the narrow annular mean background technique that uses the ULWL approach shows a very consistent trendline graph for the four background ratios used. By circular mean background technique, the trendline for the background ratio of 8:1 and 6:1 does not coincide with the trendline of the background ratio of 12:1 and 10:1. While using the FWHM technique, the ratio factor is totally inverse with the threshold value.

Close

Fig 10: Ratio factor versus threshold value using circle mean background counts techniques.

Export to PPT

Close

Fig 11: Ratio factor versus threshold value using narrow annular with ULWL applied techniques. ULWL: Upper lower window level

Export to PPT

Close

Fig 12: Ratio factor versus threshold value using FWHM techniques. FWHM: Full wave with half maximum

Export to PPT

Thus, the threshold value on the gradient different weight threshold method basically refers to the background ratio calculation, where the maximum voxel value in the targeted tumour is divided by the mean background counts. According to Kulterer et al., for every single volume that wants to be segmented, we do not know the TBR.^[18,19] Hence, the graphs in Fig 10-12 were plotted to identify the ideal threshold value for the gradient different weight threshold method. The trendline plotted in the graphs of Fig 10 and 11 fulfils the law of decay of the half-life pattern mentioned by Ramonaheng et al.,^[20] while Fig 12 shows the vice versa.

Out of the three graphs plotted, the FWHM technique's graph has inverted from the original law. As is known, the FWHM is calculated as the diameter of the volume using this technique. The FWHM concept is very effective only when a volume has a maximum voxel count that is significantly different from the mean background counts. This will result in a bell curve from the volume that has total counts higher than the mean background counts. Then, the curve seems to be very tapered, and the FWHM value will be almost the same as the actual diameter of the volume. However, when the maximum voxel count in volume is almost equal to the mean background count, the line profile needs to be drawn longer to obtain the FWHM value and further make the resulting bell curve wider. The FWHM value will be larger than the actual diameter value. Next, this will make the diameter of the volume larger than the actual value.

Although the graph threshold value using the narrow annular mean background is ideally plotted, this trendline has the lowest ratio factor, which is only 1.031, while still being used to get the threshold value. If the value of the ratio factor of a volume is lower than this value, then the threshold value will be negative. Thus, segmentation cannot be made. This will be faced by a volume that has a small volume with a low maximum count of voxels, almost the same as the mean background obtained. This will increase the segmentation volume percentage error by up to 40% or more.

The segmented sphere volume in the NEMA IQ phantom, based on the optimised threshold value, is displayed in Fig 13. Out of the six spheres employed in the gradient different weight threshold method approach, only five are searched for the threshold value. Based on the actual sphere volume, all the sphere shapes found in Fig 13b display a solid shape that follows the optimal threshold value.

Close

Fig 13: The sphere volume masked after threshold value optimisation. (a) The whole SPECT co with CT; (b) The volume spehere after thresh value; (c) Mask. SPECT: Single photon emission computed tomography

Export to PPT

Regarding the gold standard of mean background count for image quality evaluation, the background counts on the NEMA IQ phantom were determined by drawing a circular ROI on the background^[21–23] as shown in Fig 14a. However, regarding the real patient images, it is challenging to determine the normal uptake for use as background counts due to the distribution of ¹³¹I through the whole body. The uptake for the real tumour is also different for each organ in the patient's body.

Close

Fig 14: (a-c) NEMA IQ phantom for ROI background counts. ROI: Region of interest, STD: Standard deviation

Export to PPT

Furthermore, according to Li et al.^[13] If three different ROIs are drawn in the background as suggested in the image quality standard analysis, three different values of mean background counts will be obtained, even though theoretically and logically, the mean background counts should be the same. This is due to the preparation of activity concentration on the background NEMA IQ phantom being the same. This becomes worse if applied to actual patient images since the distribution of counts on image patients is uneven compared to the distribution of counts on the NEMA IQ phantom background.^[24,25]

In addition, in research proposed by Li et al., a novel background selection method is a ring-like background ROI drawing.^[13] Fig 14b shows the targeted tissue ROI drawn next to a small circular background region. This approach aggregates and eliminates the variations in counts per voxel surrounding the target tissue ROI. Additionally, this approach is simple to implement, and standardising the backdrop ROI selection may increase the reproducibility and dependability of the results. The average counts per voxel are more consistent once it has been chosen.

This method has been applied to the NEMA IQ phantom sphere for segmentation. The ROI for each sphere is displayed adjacent to the desired sphere volume, as depicted in Fig 14b. Meanwhile, Fig 14c presents an ROI of background counts that is not spherical for the targeted volume tumour with a non-spherical shape. With this approach, even if the operator draws the ROI differently, it is simpler to identify the background ROI. Although the percentage may vary amongst operators, there might be a variation.

Although the ULWL-guided narrow annular background technique was the most consistent in this study, the placement of the annular background ROI remains an important practical consideration. When a colour map is applied, it can be challenging to position a thin annulus that truly represents the local background, particularly in heterogeneous uptake. In general, background placement is more reliable when a clear intensity boundary is visible between the target and surrounding tissue (as illustrated in Fig 15a). Where the boundary is ambiguous, background estimates may be biased and can affect the derived threshold.

Close

Fig 15: Two different conditions to draw the annular narrow background. (a) Volume targeted that have high total counts compared to mean background counts; (b) Volume targeted with total counts almost equal to the mean background counts

Export to PPT

Otherwise, it is very difficult to draw a narrow annular background on the target volume whose total counts are almost equal to the average background counts because there is no clear boundary colour to distinguish between the target volume and the background, as shown in Fig 15a. However, if we look at Fig 15b, it is clear in the black dotted square area that the distribution of counts of the target volume separates between the volume and the background. Therefore, a narrow annular background is drawn outside the dotted area. This method is more skillfully used when an operator frequently practices on volumes with low total counts.

One factor that influences segmentation performance is the contrast (sphere-to-background ratio) and the available count statistics. In this study, four sphere-to-background ratios (12:1, 10:1, 8:1, and 6:1) were used to span a clinically relevant range for post-therapy I-131 SPECT, where tumour-to-background contrast is often modest and total counts are limited because imaging is performed days after administration. The selected range also overlaps with ratios used in prior phantom work (e.g., including 10:1 and 6:1).^[26] While higher contrasts can occur in selected lesions, the chosen ratios were intended to represent common low-to-moderate contrast scenarios where robust delineation is most challenging.

Proof-of-concept evaluation of gradient-different-weight threshold method with patient data

The proof-of-concept evaluation included two retrospective patient datasets. Case 1 had uptake predominantly localised to the thyroid region, while Case 2 showed I-131 distribution throughout the whole body [Fig 16].

Close

Fig 16: (a) Axial case one; (b) Sagital case one; (c) Coronal case one; (d) Axial case two; (e) Sagital case two; (f) Coronal case two

Export to PPT

Reference (ground-truth) tumour segmentations for the clinical cases were delineated independently by two nuclear medicine physicians [Table 2]. For quantitative evaluation, the reference volume was defined as the average of the two physician-derived volumes. Ethical approval for this study was granted by the Medical Research and Ethics Committee (MREC) under the Ministry of Health Malaysia (NMRR-20-1239-53306).

Inter-observer variability was summarised from the two physician segmentations in Table 2. Across 8 lesions, the absolute volume difference between observers ranged from 0.18 to 1.71 mL (mean 0.72 mL; median 0.42 mL). The relative difference (|V1−V2|/mean) ranged from 4.8% to 30.5% (median 7.0%). This uncertainty is larger for small lesions and should be considered when interpreting DSC and volume error results.

The segmentation results of the proposed method are quantitatively evaluated in this experiment using metrics commonly used in medical image segmentation. The percentage error (%), RMSD and DSC are the corresponding evaluation matrices.^[27–29]

All the volume tumour segmentation results for both cases are shown in Table 3-5 that display the three 3D images for both patient cases using the segmentation methods studied. 3D image display is a registration process between SPECT and CT imaging data sets.

Table 3: Comparison of two case patient data for gradient different weight method verification.

Case	Method	3D Segmentation
Gradient different weight	Case number one
	Case number two

When the gradient different weight threshold method was applied to case number one, only one tumour was visible on the thyroid organ with a volume of 27.1 mL. This method required determining a mean background count to obtain the threshold value. Hence, the mean background counts were taken around the tumour using the narrow annular mean background technique.

^aCalculated as the average volume segmentation between two nuclear medicine physicians.

For case number two, eight visible tumours were identified and agreed upon by the two nuclear medicine physicians. [Table 5] compares four segmentation methods for measuring tumours in case number two.

This situation is challenging compared to case number one, which is the tumour segmented on the thyroid region is not connected with the background counts. Case number two has many tumours all over the body, especially in the liver region. When a tumour to be segmented is specified, the tumour's volume will merge with the surrounding area counts. Hence, the volume calculation of a tumour is irrelevant. This method is not suitable for the image of patients with a large distribution of tumours and low count values.

The gradient different weight threshold method in case number two identified eight completely predefined tumours, as shown in Table 5. The tumours were 8.69, 14.36, 8.11, 21.87, 27.78, 6.67, and 4.39 mL in volume. The mean background counts were taken adjacent to their respective tumours using the narrow annular mean background counts technique. In the last evaluation method of 3D U-Net deep learning, the segmented tumour volume appears more and relatively larger than the tumour volume obtained by the gradient different weight threshold method. Each volume was 31.1, 29.4, 44.3, 7.44, 16.51, 17.32 and 4.1 mL.

Typically, for tumour volume segmentation on patient images, the nuclear medicine physician often employs the fixed threshold method as the primary technique for segmentation. The fixed threshold method is preferred for its ease of determining tumour volume and speed. However, an unresolved issue remains regarding the percentage of the fixed threshold method's threshold used from the maximum voxel value. The execution of these tasks largely depends on the operator's experience. The choice of percentage for the maximum voxel also relies solely on the visual interpretation of the image and the perceived accuracy of the tumour volume observed. For instance, the thresholds for tumour volume segmentation of ¹⁷⁷Lu and ⁹⁰Y differ; typically, the threshold for ¹⁷⁷Lu is between 40-50%, while for ⁹⁰Y, it can go up to 70% of the maximum voxel value. This is due to variations in the image quality, specifically concerning the contrast ratio.

Nevertheless, the fixed threshold method is not accurate. In fact, there are many sources that say the fixed threshold method is not suitable for tumour segmentation for SPECT images, except for images using ^90mTc as a radiotracer. The main factor is the partial volume effect (PVE). The term “PVE” is frequently employed in nuclear medicine imaging to refer to the loss of contrast between an object and its surroundings. Alternatively, the concept of PVE can be described as the interplay between spill-out (assigning object counts to the background) and spill-in (attributing background counts to the object). In a hot area, this impact may lead to a decrease in absolute intake. PVE is often mentioned in relation to spatial resolution, which measures how sharp or in-depth an image is.

Among the methods explored in this study, the gradient-different-weight threshold approach showed encouraging agreement with physician-derived reference segmentations in the proof-of-concept cases. However, analysis of the retrospective data highlighted an important limitation: in addition to tumour-to-background ratio and tumour geometry, inter-lesion proximity can bias the local background estimate. When lesions are close together, an annular background ROI drawn around lesion 2 may include counts from lesion 1, inflating the estimated mean background and altering the resulting threshold [Fig 17]. When lesions are well separated, the annular ROI is more likely to represent the true local background.

Close

Fig 17: The distance of two tumours in case number two.

Export to PPT

According to previous studies, the tumour’s size and shape are among the many variables that affect how the recovery curves are shaped, as well as the radioactive source that is present around the object. In this study, the threshold value trendline curves were determined using solely the sphere volume set up in water with a radioactive source of the NEMA IQ phantom. The outcome could be incorrect when these are for structures with irregular shapes. In this light, it should be noted that the geometry of the determination of the threshold value will be influenced by tumour volume.

DISCUSSION

This study developed a modified gradient-different-weight threshold approach for tumour delineation in I-131 SPECT and performed phantom-based calibration followed by a proof-of-concept clinical evaluation. The workflow combines (i) threshold optimisation using known sphere volumes in a NEMA IQ phantom and (ii) a ratio factor derived from the maximum voxel count within the target and the estimated mean background, providing a practical link between lesionto-background characteristics and the threshold used for segmentation.

A key finding is that the narrow annular mean background technique with ULWL produced the most consistent ratio-factor–versus–threshold relationship across the tested sphere-to-background ratios, compared with circular background ROIs and the FWHM-based approach.

This is an important practical result because threshold-based delineation often fails in SPECT due to background nonuniformity and operator-dependent ROI selection. In the phantom, the circular ROI background approach showed less consistent behaviour at lower contrast ratios, suggesting greater sensitivity to local background heterogeneity and ROI placement, while the ULWL-guided annular ring approach better “standardised” the local background sampling around each target.

The FWHM-derived background technique produced an inverted relationship between the ratio factor and the threshold. Mechanistically, this behaviour is expected when Cmax approaches Cmean: the line-profile becomes broad, and the estimated FWHM expands beyond the true object diameter, inflating the region used to define the boundary/background and ultimately distorting the threshold selection. This reinforces that FWHM-based boundary definitions can be reliable primarily when lesion contrast is high, and the peak-to-background separation is clear.

Although the ULWL annular method produced the most “well-behaved” trendline, it also yielded the lowest ratio factor floor (~1.031); below this, the model can produce negative threshold values, making segmentation impossible and increasing volume error substantially (reported up to ~40% or more in this scenario).

This limitation is clinically relevant in I-131 SPECT, where lesions may show weak uptake, and the surrounding activity distribution can be complex. In practice, this suggests the method should be applied with a minimum contrast/ratio-factor criterion, or augmented with a fallback strategy (e.g., clinician-guided seed constraints, CT-guided masks, or adaptive regularisation) when the local contrast is too low.

The two retrospective cases were intentionally different: Case 1 had uptake mainly localised to the thyroid region, while Case 2 demonstrated more widespread I-131 distribution; ground truth volumes were provided by two nuclear medicine physicians (MREC/NMRR approval reported).

In case 1, the proposed method showed close agreement with the physician-derived reference volume (average), with DSC ≈ 0.9541 and a modest volume error.

Case 2 highlighted a key clinical limitation: when lesions are numerous and/or surrounded by heterogeneous activity, the target can merge with surrounding counts, making the segmented volume unreliable if the background estimate is contaminated by nearby uptake.

The study also identified that inter-lesion distance can bias the annular background (the annulus of one lesion may include counts from an adjacent lesion), effectively raising the “background” and altering the resulting threshold.

This finding is clinically important because metastatic disease often contains clustered lesions, particularly in organs such as the liver, where spill-in/spill-out and partial volume effects further complicate delineation.

The manuscript notes that fixed-percentage thresholding is widely used clinically due to its simplicity and speed, but the choice of percentage is strongly operator- and tracer-dependent and is influenced by image contrast and partial volume effects.

A limitation of this work is that a full quantitative, side-by-side benchmarking against alternative segmentation approaches (e.g., fixed-percentage thresholding with optimised percentage, CT-assisted methods, or other contemporary algorithms) using identical datasets and evaluation metrics was not performed. Future work should include direct comparisons on the same phantom and multi-patient datasets to more objectively quantify any added value over current clinical practice.

The proposed method addresses this by providing a systematic threshold selection framework that is linked to a measurable local ratio factor and calibrated using phantom data. However, because the calibration was derived from spherical volumes in uniform phantom conditions, transfer to irregular tumour shapes may introduce bias; thus, tumour geometry is expected to influence the threshold–ratio relationship.

Future improvements should focus on: (i) improving background selection to avoid contamination from adjacent lesions (e.g., exclusion masks, multi-ring sampling, or robust statistics), (ii) extending calibration beyond spheres to include irregular inserts and lower contrast regimes, and (iii) larger multi-patient validation with stratification by lesion size, contrast, organ site, and lesion clustering to define a clear clinical operating range for the method.

CONCLUSION

A modified gradient, different-weight threshold segmentation framework was developed for SPECT I-131 tumour delineation, using phantom-derived threshold optimisation and a ratio-factor concept to reduce the operator dependence typical of fixed-threshold methods.

Among the evaluated background estimation strategies, the ULWL-guided narrow annular background produced the most consistent ratio-factor–versus–threshold relationship across multiple sphere-to-background ratios, supporting its use as the preferred background approach for this workflow.

In a proof-of-concept clinical evaluation limited to two retrospective cases, the method achieved high agreement with physician-derived reference segmentation in a case with predominantly localised uptake (DSC ≈ 0.9541), demonstrating potential for practical adoption in I-131 SPECT-based tumour contouring for dosimetry workflows.

However, performance can degrade in disseminated disease or low-contrast settings where lesions are close together, and background estimation becomes contaminated, and a lower ratio-factor boundary may lead to failure (negative threshold) for very low-contrast lesions.

Because the clinical evaluation included only two cases, the results should be interpreted as preliminary. Larger retrospective and prospective cohorts, with reporting of inter-observer variability and lesion-level stratification (size, contrast, and clustering), are required to establish generalisable performance.

Overall, the proposed approach offers a structured and reproducible pathway for I-131 SPECT tumour delineation, with clear directions for refinement, particularly for clustered lesions, heterogeneous backgrounds, and irregular tumour geometries.