Сегментация с искусственным интеллектом и автоматизированный расчет индекса каллосу с корпусом при множественном склерозе.

M ultiple sclerosis (MS), which develops through autoimmune mechanisms triggered by environmental factors in genetically sensitive individuals, is a chronic disease characterized by central nervous system demyelination. The number of people estimated to have MS worldwide reached 2.8 million in 2020, and women are 2-3 times more likely to develop the disease. 1 It is generally seen in young adults and the most common form has a course of attacks and remissions. 2

Due to the accumulated processes of chronic inflammation, demyelination, and axonal damage in MS patients, atrophy and loss of volume occur in structures such as the neural parenchyma, grey and white matter, and the corpus callosum. 3 , 4 The corpus callosum has been shown to atrophy more than other brain structures and could represent a more sensitive marker than other measures of brain atrophy and a strong predictor of clinical disability. 5

The basic principle of MS diagnosis is the observation of lesions in the central nervous system and the resulting clinical condition spreading over time, and distinguishing it from other disorders with overlapping features through precise clinical assessment and appropriate diagnostic techniques. 6 The most important parameter in current practice, after clinical evaluation for both diagnosis and disease follow-up is magnetic resonance imaging (MRI). Characteristically, linear-ovoid-shaped demyelination plaque appearing hypointense on T1-weighted (T1W) images and hyperintense on T2-weighted (T2W) images, and contrast-enhancing lesions defined as active plaque are the routinely used MRI parameters. 7 With MRI examination, the periventricular area, floor and roof of the 4th ventricle, periaqueduct, pons, and optic nerve spinal cord, where demyelinating plaques are seen intensely can be shown, as well as demyelination or atrophy of the corpus callosum. 8 In addition to traditional methods, artificial intelligence (AI) techniques used today enhance the detection of MS-specific lesion features by recognizing visual patterns, identifying textures beyond human perception in MRI sequences, and automating time-consuming tasks such as segmentation. As a result, challenges like subjective interpretation by clinicians and interobserver variability can be minimized. 9

Atrophy of the corpus callosum provides a measure of the long-term accumulated effects of MS. Research suggests a correlation between corpus callosum atrophy and the duration and/or intensity of the disease, with disruptions in its microstructural integrity being linked to cognitive impairments, including deficits in information processing speed, visual and verbal memory, and executive functions. 10 , 11 Furthermore, corpus callosum atrophy has been shown to be a substantial predictor of clinical disability.

There are imaging-based techniques in the literature that recommend manual 2D measurements of the corpus callosum index (CCI) and corpus callosum area or volumetric measurements with volumetric software methods for assessing the corpus callosum in MS patients. 3 , 7 , 12 - 14 Compared to the volumetry technique, CCI is an easy measurement technique, recognized for its good inter-rater, intra-rater, and inter-scanner reliability. 3

On conventional MRI, the CCI is calculated by dividing the total thicknesses of the corpus callosum genu, corpus, and splenium by the anterior-posterior length of the corpus callosum, which is considered a practical approach for evaluating atrophy in chronic MS patients. 15 There are also recent studies that have been conducted of MS diagnosis with AI techniques in the segmentation of plaques and the measurement of corpus callosum volume, thalamus volume, and segmentation. 16 - 18 Artificial intelligence-based tools can perform corpus callosum segmentation with high accuracy and concordance rates. However, performance consistency across different sequences and MRI scanners, as well as the integration of AI models into clinical workflows, have not yet been achieved at sufficient levels. This study aimed to investigate the differences in CCI between MS patients in the chronic phase and healthy individuals, and to evaluate corpus callosum segmentation in MS patients using AI techniques.

Methods

This retrospective study was conducted in Akdeniz University Faculty of Medicine, Antalya, Turkey with the approval of the Local Ethics Committee. Informed consent was not required from the study participants as the study was planned to be retrospective. Between January 2019 and August 2021, cranial MRI examinations of a total of 300 cases in the patient and control groups were screened. The study group consisted of 150 patients aged 20 to 50 who had been diagnosed with different subtypes of MS and were regularly monitored by an experienced physician in the Department of Neurology. Exclusion criteria were patients with neuromyelitis optica spectrum disorders, any brain pathology other than MS, and patients with motion artifacts on their MRI that affected the CCI measurement. The control group was formed of 150 healthy cases, aged 20-40 years, who had undergone cranial MRI in our unit for another initial diagnosis and had no additional disorders.

Imaging parameters and analysis

The images used in the examinations were acquired with 1.5 Tesla and 3 Tesla MRI scanners (Magnetom Avanto and Spectra, Siemens Healthcare, Erlangen, Germany). In the routine MRI of MS patients, T1W axial, T2W axial, T2-FLAIR axial and sagittal, susceptibility-weighted imaging and diffusion-weighted MRI sequences were obtained. On the sagittal FLAIR sequence to be used for the measurements, the TE value was determined as 336 msn, TR value as 5000 msn, and slice thickness as 1.4 mm. In the routine MRI of the control group subjects, T1W sagittal, T2W axial, T2-FLAIR axial and sagittal, susceptibility-weighted imaging and diffusion-weighted MRI sequences were obtained. On the T1W sagittal sequence used for the measurements, the TE value was determined as 8.7 msn, TR value as 490 msn, and slice thickness as 5 mm.

The CCI measurements were performed by a single radiologist using the method described by Figueira et al. 15 From the mid-sagittal slices of the corpus callosum which could be most clearly selected, the anterior section thickness, mid-section thickness, and posterior section thickness values of the corpus callosum were totalled, and then divided by the anterior-posterior long axis value of the corpus callosum and thus the CCI value was obtained ( Figure 1 ). These values were recorded and compared with the data determined with AI techniques.

- Manual corpus callosum index measurements of a study group case.

Corpus callosum segmentation with AI techniques

Images in JPEG format were obtained from the most clearly differentiated slice of the corpus callosum on the MRI of the study group patients and the control group subjects. As the corpus callosum region was determined with the help of computerised imaging methods, the necessary limit information was obtained by software engineers experienced in the field before the CCI value measurement.

For model training in the supervised learning method, the OASIS and ABIDE references were used as the dataset for labelling. The data for healthy individuals is held in OASIS, and the data for MS patients in ABIDE. It was aimed to define the datasets of the anatomic region by labelling the anterior section, corpus, and posterior section of the corpus callosum using the aforementioned datasets. Using data collected from the 2 sources, a dataset comprising approximately 2000 labeled images was created. To address the differences between the OASIS and ABIDE databases, harmonization steps such as spatial standardization, intensity normalization, and resolution standardization were applied. Additionally, images with motion artifacts or low signal-to-noise ratio were excluded from the study.

Explanation of the model architecture

Python language, which is often used in deep learning applications, was used as the software language for this study. Code blockers developed for the model training and application processes were developed with Python. The Tensorflow and Keras modules, which are compatible with Python software language, were used to define the model architecture and for training.

The U-Net model, known for its high effectiveness in medical applications, was selected as the model architecture for this study due to its faster convergence in terms of training time ( Figure 2 ). Each blue box in the figure represents a multichannel attribute map. The channel number is above the box, and the size information is in the lower left corners. The white boxes are duplicate attribute maps. The processes applied are indicated with different arrows.

- The architecture of the U-Net model. In the image, each blue box represents a multi-channel feature map. The number of channels is at the top of the box. The size information is in the lower left corners. The white boxes are copy feature maps. The applied operations are indicated with different arrows.

i. Conv 3x3, ReLU: This is the convolution process performed with a 3x3 mask. Masks were made with the Rectified Linear Unit activation function. If the value entering this function is positive, the output is given in the same form. When the entry is negative, a value of 0 is given in the output.

ii. Copy and crop: This is the process to copy and crop the attribute map.

Max pool 2x2: This is the function in which the largest value applied with a 2x2 mask is taken as the output.

iii. Up-conv 2x2: This is used for conversion from low resolution to high resolution. The resolution of the pixel value is increased with the parameters learned in training. It is performed with 2 x 2 masks.

iv. Conv 1x1: This is the convolution process applied with a 1x1 mask.

Training stages, model metrics, and automatic CCI calculation

To prepare the dataset for model training, the Numpy module, compatible with Python, was used. With the use of this module, all the images in the dataset and the segmented mask labels were made ready for model entry. Due to the architectural structure, the width and height were adjusted to be 128 pixels by re-sizing on the images. The series formed following the re-sizing was recorded in “pkl” format. Of the set formed, 70% was used for training, 10% for validation, and 20% for testing.

The recording and reading processes of the datasets processed in “pkl” format were performed with the Joblib module. In addition, the data were enriched by applying rotation processes on the dataset.

By reloading all the marked dataset information (marked images) obtained into the Tensorflow and Keras modules at the training stages of the AI modality, it was aimed to differentiate the difference in the anatomic structure of the corpus callosum from other brain regions, and by the anatomic removal of this area to obtain it alone.

Before training, initial adjustments were made by obtaining the mean and standard deviation values of the pixel values of the whole dataset. The model training metrics were set as Smooth: 1.0, Dropout rate: 0.5, Epoch number: 50, Batch size: 32, Kernel size: 5x5, and initial volume size: 64.

The main aim of the solution presented in the scope of this study was to automate the calculation of the CCI. By applying the specifically coded CCI calculation algorithm to the segmentation outputs obtained in the first stage completed with the segmentation process, the index value was obtained.

Statistical analysis

Statistical analyses were performed using IBM SPSS version 25. Descriptive statistics were presented as mean±SD and min-max values for continuous variables and as n(%) for categorical variables. Normal distribution of CCI measurements was confirmed via Kolmogorov-Smirnov test and Skewness-Kurtosis tests, enabling parametric testing. Independent Samples t-test compared means between study and control groups, while ANOVA evaluated differences across MS subtypes. ROC analysis determined CCI cutoff values with corresponding AUC, sensitivity, and specificity. Pearson correlation coefficients assessed relationships between continuous variables, while Chi-square test examined associations between categorical variables. Statistical significance was defined as p <0.05.

Results

The study was conducted between January 2019 and August 2020. The patient group of 150 MS patients in the chronic period comprised 49 (32.7%) males and 101 (67.3%) females with a mean age of 40.7±8.2 years (range, 20-50 years). The control group of 150 healthy subjects comprised 70 (46.7%) males and 80 (53.3%) females with a mean age of 29.3±5.5 years (range, 20-40 years). The demographic data are presented in Table 1 .

Table 1 - Demographic data of the participants in the study. Variables Patient group (n=150) Control group (n=150) P -value Age (years) (mean±SD) 40.7±8.2 29.3±5.5 0.001 Gender (%)


0.018 Male 49 (32.7) 70 (46.7)
Female 101 (67.3) 80 (53.3)
Course of disease (%)
RRMS 119 (79.3)

SPMS 20 (13.3)

PPMS 11 (7.3)

Duration since first manifestation of disease (years) (mean±SD) 12.6±5.9

RRMS 11.6±5.4

SPMS 17.4±5.1

PPMS 15.7±7.4

Corpus callosum index (mean± SD) 0.274±0.048 0.382 ±0.030 0.01 RRMS: relapsing-remitting multiple sclerosis, SPMS: secondary progressive multiple sclerosis, PPMS: primary progressive multiple sclerosis, SD: standard deviation

- Demographic data of the participants in the study.

RRMS: relapsing-remitting multiple sclerosis, SPMS: secondary progressive multiple sclerosis, PPMS: primary progressive multiple sclerosis, SD: standard deviation

The diagnoses of the cases in the study group were 79.3% relapsing-remitting MS (RRMS), 13.3% secondary progressive MS (SPMS), and 7.3% primary progressive MS (PPMS). A statistically significant difference was determined between the MS subtypes in respect of age, time since diagnosis, and CCI score. The mean CCI was calculated as 0.281 in the RRMS group, 0.240 in the SPMS group, and 0.246 in the PPMS group. In the RRMS group of patients, the CCI score was determined to be statistically significantly higher, and the time since diagnosis and patient age were statistically significantly lower compared to the other subtypes ( p₁ =0.01; p 2 =0.01; p 3 =0.02).

The CCI score was calculated as mean 0.274 in the patient group and 0.382 in the control group. The difference between the groups was statistically significant, with a higher index value determined in the healthy control subjects than in the MS patients ( p =0.01). The distribution of CCI scores according to gender reported in studies in literature is shown in Table 2 .13,19-23

Table 2 - Comparisons with literature of the distribution of the mean corpus callosum index (CCI) values in the patient and control groups. Authors Number of cases age, mean (years) MRI sequence Slice thickness Disease subtype (n) RR/SP/PP Disease duration, years (mean) Mean CCI values Patient group Control group Patient group Control group Perez-Alvarez et al (2018) 109 37.8 101 37.7 T1 (Sag) 3 mm 90/0/7 *
- 0.377 0.411 Granberg et al (2015) 23 57 23 57 T1 (Sag) T2 FLAIR(Sag) 4 mm 3/20/0 27 0.274 0.393 Cilingir et al (2017) 212 35.1 52 35.6 T2 FLAIR(Sag) 2 mm 177/20/15 7.1 0.341 0.386 Fujimori et al (2020) 75 40.3 21 36.2 T1 (Sag) T2 FLAIR(Sag) 3 mm 1.5 mm 59/12/1 9.6 0.370 0.420 Schependom et al (2017) 29 48 15 46.6 T1 (Sag) 2.5 mm 29/0/0 17 0.319 0.362 Nishizawa et al (2022) 91 41.6 41 40.1 T1 (Sag) T2 FLAIR 1.4 mm 71/14/2 **
9.13 0.375 0.444 The current study 150 40.7 150 29.3 T1 (Sag) T2 FLAIR(Sag) 5 mm 1.4 mm 119/20/11 12.6 0.274 0.382 * The diagnosis of 12 patients in the study was isolated demyelinating syndromes. ** The diagnosis of 4 patients in the study was clinically isolated syndrome, Sag: Sagittal, RR: relapsing-remitting, SP: secondary progressive, PP: primary progressive

- Comparisons with literature of the distribution of the mean corpus callosum index (CCI) values in the patient and control groups.

The diagnosis of 12 patients in the study was isolated demyelinating syndromes.

The diagnosis of 4 patients in the study was clinically isolated syndrome, Sag: Sagittal, RR: relapsing-remitting, SP: secondary progressive, PP: primary progressive

According to the ROC analysis in this study, a CCI cutoff value of 0.334 showed 98.3% differentiation between the patient and control groups. Sensitivity and specificity were calculated as 94%, showing the cutoff value to have good differentiation with high accuracy.

No statistically significant correlation was determined between the CCI score and patient age or gender in both the patient and control groups ( p 1 =0,812; p 2 =0,924). The results of the correlation analysis of the measurements in the patient group showed a statistically significant negative correlation between the CCI score and the time since diagnosis (r:-0.837, p =0.001). As the time since diagnosis increased, the CCI score decreased.

The result of the training dataset used with the AI techniques reached a success point of 0.9381 in the patient group. The training process and clear anatomic differentiation of the corpus callosum from other structures were successful in the MS patient group. The mean CCI calculated automatically after segmentation in the study group was 0.286. Although the mean CCI scores measured manually and automatically were extremely close to each other, a statistically significant difference was determined ( p =0.04). No statistically significant difference was determined in the CCI score calculated with automatic segmentation according to gender and MS subtype ( p₁ =0.58, p 2 =0.058).

When the MRI of the control group, taken with the appropriate protocols and labelled, which were not included in the training set, were shown to this described model, the desired results could not be obtained. Even if the corpus callosum region was correctly determined in the control group, it was seen that some areas other than the corpus callosum were included in the outputs of the training model. As the AI modelling techniques used in this study gave false results, they were not applied to the control group subjects.

The original MRI of the cases in the patient and control groups, the corpus callosum region, and the model prediction outputs are shown in Figures 3 & 4 .

- In the study group patients, the original versions of the images, the corpus callosum region, and, the model prediction outputs are seen.

- Original images and artificial intelligence outputs of control group patients. In predictions made for the control group protocol, the corpus callosum region was generally identified correctly. However, it was observed that some adjacent areas of the corpus callosum that are not part of it were also included in the segmentation.

Discussion

In this study, the CCI was measured manually on the MRI of a healthy control group and a study group formed of patients with different subtypes of MS. Using the index values obtained it was attempted to determine the presence of corpus callosum atrophy in the MS patients and to investigate the relationship of this atrophy with the factors of MS subtype, age, gender, and time since diagnosis. The mean CCI was 0.273 in the patient group, which was significantly lower than the value in the control group. Similar to studies in the literature, a cutoff value of 0.334 was determined in the CCI measurement with 98.3% differentiation between the patient and control groups at sensitivity and specificity level of 94%. 19 , 23

Corpus callosum atrophy is directly associated with disease progression, with pronounced atrophy particularly observed in patients with secondary progressive multiple sclerosis SPMS. Corpus callosum atrophy has been linked to impairments in upper and lower extremity motor functions, deficits in balance, and fine motor skills among MS patients. Moreover, a reduced CCI correlates with cognitive dysfunction, manifesting as impairments in memory, attention, executive functions, and processing speed. 5 , 11 Therefore, in MS patients, CCI can serve as a neuroimaging biomarker for assessing disease progression and predicting cognitive and motor function decline. When considering other widely used atrophy indices, for example, the brain parenchymal fraction (BPF) is an automated segmentation tool that measures changes in total brain volume. It has been shown to strongly correlate with disability progression in MS and is primarily used for monitoring disease progression. 24 However, BPF may not capture regional atrophy as effectively as the CCI, which specifically reflects the integrity of the corpus callosum, a structure frequently affected in MS. Similarly, thalamic atrophy has emerged as a predictor of cognitive impairment and disability in MS. 25 While thalamic atrophy provides valuable insights into subcortical pathology, CCI offers a more focused assessment of white matter integrity. However, the limitations of CCI may become apparent in cases where atrophy is predominantly cortical or infratentorial.

In a retrospective study of 30 MS patients, Sugijono et al 3 found that the brain volume and CCI of MS patients were significantly smaller than those of healthy subjects. A significant correlation was found between the CCI and corpus callosum volume and cerebral white matter volume, and between the CCI and cortical and subcortical grey matter. 3 In another study of 235 MS patients, the CCI was found to be related to the total brain volume obtained with volumetric techniques, and it was concluded that the CCI could reflect the presence of atrophy in the early stages of demyelinating diseases, making it a rapid and easy alternative calculation tool for this. 20

Eser et al 26 conducted a study aiming to determine the corpus callosum normal measurement values in healthy individuals, and reported that there was a statistically significant difference in favor of males in respect of the corpus callosum area, genu, and anterior-posterior diameters, but no significant difference in the truncus and splenium measurements. In the current study, the CCI score was found to be similar in males and females in both the patient and control groups.

When MS subtypes were examined, Sugijono et al 3 reported that the CCI was higher in the RRMS group compared to the other subtypes, which was similarly found in the current study. Figueira et al 15 determined a negative correlation between the CCI and the time since diagnosis, which was also similar to the current study. This finding demonstrates that as the duration of MS increases and the disease becomes chronic, so the atrophy of the corpus callosum progresses.

In the routine cranial MRI protocol in our clinic, slice thickness is 5 mm, and T2W FLAIR axial slices are taken but not T2 FLAIR sagittal plane. As this study was retrospective in design, the CCI was measured on T2W FLAIR sagittal 1mm thickness slices in the MS patient group and on T1W sagittal 5 mm thickness slices in the control group. In a study by Perez-Alvares et al, 20 CCI measurements were taken on different sequences and different slice thicknesses (0.8-5 mm) of an MS patient group and a healthy control group, and a significant difference was determined between the groups. There are also other MRI studies in the literature which have measured the CCI on different sequences and slice thicknesses. 3 , 13 , 21

Manual measurement of the corpus callosum is accepted as the best standard method to measure callosal atrophy. 13 However, it is not known how accurately these measurements can be repeated between different MRI scanners and the readers in different centres. With the rapid development and use of AI and automatic segmentation methods, more accurate and less time-consuming alternative measurements can be provided independently of the user. Machine learning can differentiate between normal and pathological brain tissues to detect atrophy, while deep learning and segmentation algorithms enable the automatic measurement of the CCI and volume from MRI. Consequently, automated image processing offers a more accurate, objective, and effective approach for managing multiple sclerosis with early diagnosis and personalized treatment plans. Using AI techniques in the current study, it was aimed to achieve corpus callosum segmentation and automatic CCI calculation. Successful automatic segmentation was able to be made in the MS patient group. Although the CCI mean values were calculated to be statistically significantly different from the manual measurements, the mean values were similar and when significantly lower than the cutoff value of 0.334, reflected corpus callosum atrophy. However, while manual measurements revealed a significantly higher CCI score in the RRMS group compared to other MS subtypes, this difference was not detected in automated measurements. This discrepancy is thought to be due to the considerably lower number of patients in other subtypes compared to the RRMS group in our study. Future studies with larger case series, using homogenous distribution of subtypes, are necessary to validate and correlate these findings.

The U-Net model’s ability to be effectively trained using relatively small datasets, its capability to directly process images obtained from standard imaging devices, and its capacity to perform segmentation in a simple and rapid manner compared to manual methods offer significant advantages in routine clinical practice. The U-Net architecture in this study was also seen to provide quite successful results for the patient group. For control group protocols, the preparation of labelled datasets and training made with the same hyperparameters on the same architecture will be able to obtain models with high success rates. In this way, by performing automatic segmentation in other protocols used in daily routine in different clinics, the pre-calculation processes of the CCI value can be automated.

Platten et al 18 compared manual and automatic FreeSurfer segmentation methods to calculate the corpus callosum area on the MRI of 9 MS patients taken on different MRI devices. The FreeSurfer longitudinal stream method was found to be more sensitive than the gold standard manual methods measuring the area, but due to limited accuracy the selection of the manual definitions was recommended for the separation of the corpus callosum into segments. In the current study, there was successful clear anatomic differentiation of the corpus callosum from other structures in MS patients.

Study limitations

This study had both strengths and limitations. One of the most important limitations is that the MRI obtained for the patients and the control group were different sequences and of different thicknesses. Due to the differences in image quality formed by these sequences, the CCI was measured more successfully in the patient group from higher resolution, thin-section images. In the patient group, AI techniques in the segmentation outputs on the sample taken with the sequences taken with the MS protocol provided results with nearly complete accuracy. However, the images of the control group were not fully compatible with the AI dataset base and training models and the corpus callosum segmentation did not reach a sufficient level. Therefore, automatic CCI values could not be obtained for the control group and so the measurements could not be compared between the 2 groups. Although the use of different sequences for the study and control groups is not ideal, it reflects what is often used in typical clinical practice as many radiology groups often use different brain MRI protocols for different indications. Future prospective studies using thin-section imaging and the same MRI sequences will help standardize image acquisition.

Another limitation could be said to be that there was no inter and intra-observer agreement evaluation as the manual CCI measurements were performed by a single radiologist. However, to minimize potential bias, all manual measurements were performed following a standardized protocol, and the methodology was validated by an experienced radiologist (KK). Despite the statistically significant difference between manual and AI-based CCI measurements ( p =0.040), the absolute difference was minimal and remained well below our predefined atrophy cut-off value of 0.334. This suggests that there is overall consistency between the 2 methods, and the minimal discrepancy does not appear to substantially impact the clinical interpretation of the results. Future studies should incorporate multiple raters and reproducibility analyses to further validate the reliability of manual CCI measurements. On the other hand, as the AI dataset base and the number of images which can be repeatedly learned in the training models are increased, AI may be able to determine the corpus callosum and size changes in this structure at higher accuracy. Additionally, AI models can be retrained by incorporating different imaging techniques or transferring various datasets. By developing image preprocessing/transformation algorithms, and optimizing the hyperparameters, the model can be adapted to different imaging protocols. Thus, AI could potentially be used reliably as a replacement for manual measurements. In the MS group, the number of patients who have been followed by the same clinician for a long period is limited. As a result, the inability to achieve a homogeneous distribution across subgroups, such as gender or time since diagnosis, along with the relatively small number of cases in the study, can be considered additional limitations.

In conclusion, the results of this study demonstrated that the CCI is a method with high sensitivity and specificity in respect of determining corpus callosum atrophy in patients with MS in the chronic phase compared with healthy individuals. In addition, the findings obtained using AI techniques such as segmentation, machine learning, and deep learning to determine corpus callosum atrophy were seen to be successful in the corpus callosum region in MS patients and the automatically calculated CCI score was successful in showing atrophy. By developing the training models and dataset infrastructure appropriate to the sequences obtained from control group subjects, corpus callosum size measurements could be made rapidly and practically in healthy individuals. The current study is the first in this area in literature and it can be considered that it will be of guidance for future studies.

Acknowledgment

We would like to thank Caroline Walker ( carolinejwalker@yahoo.co.uk ) for the English language editing.

Footnotes

Disclosure. Authors have no conflict of interests, and the work was not supported or funded by any drug company.

Disclosure. Authors have no conflict of interests, and the work was not supported or funded by any drug company.

References