Versatile Mass Spectrometry-Based Intraoperative Diagnosis of Liver Tumor in a Multiethnic Cohort

Abstract

Currently used techniques for intraoperative assessment of tumor resection margins are time-consuming and laborious and, more importantly, lack specificity. Moreover, pathological diagnosis during surgery does not often give a clear outcome. Recent advances in mass spectrometry (MS) and instrumentation have made it possible to obtain detailed molecular information from tissue specimens in real-time, with minimal sample pre-treatment. Probe Electro Spray Ionization MS (PESI-MS), combined with artificial intelligence (AI), has demonstrated its effectiveness in distinguishing liver cancer tissues from healthy tissues in a large Italian population group. As the MS profile can reflect the patient’s ethnicity, dietary habits, or particular operating room procedures, the AI algorithm must be well trained to distinguish different groups. We used a large dataset composed of liver tumor and healthy specimens, from the Italian and Japanese populations, to develop a versatile algorithm free from ethnic bias. The system can classify tissues with discrepancies <5% from the pathologist’s diagnosis. These results demonstrate the potential of the PESI-MS system to distinguish tumor from surrounding non-tumor tissues in patients, with minimal bias from race/ethnicity or etiological characteristics or operating room procedures.

1. Introduction

Surgery is the cardinal procedure to treat solid tumors, with different approaches employed for the different histological variants. Together with transplantation, tumors resection is a mainstay for hepatocellular carcinoma (HCC) treatment. Overall five-year survival after surgical resection ranges between 20 and 60%, depending on the liver function and histological profile [1].

The goal of surgery is to remove the whole tumors, with minimal loss of surrounding healthy tissue: this is essential to preserve the function of the target organ, without over-resecting the surrounding non-tumor tissue. Besides the higher healthcare costs, the risk of relapse would implicate physical and psychological morbidity for patients.

Several approaches, such as ultrasound 3D navigation and indigo carmine staining of tissue by infusion, now permit more accurate identification of the resection border during surgery. However, because the tissue is not examined directly, the diagnosis must be considered clinical rather than definitive.

On the other hand, rapid intraoperative pathology diagnosis usually fails to yield a clear and unequivocal result, leaving surgeons hesitant about a definitive decision. Furthermore, depending on the number of specimens, surgeons must hold up the surgical procedure for 15–30 min, leaving the patient under general anesthetic, which may be a problem for some patients, as well as being costly for a social medical care system.

A more sensitive and quicker procedure is therefore required to improve the surgical treatment of solid tumors. The majority of current procedures, regardless of the instrumentation repertoire, are based on morphological alterations. As a result, completely alternative approaches based on the chemical fingerprinting of the tumors may interest surgeons because they provide an entirely new layer of information based on arguably more objective data.

The use of mass spectrometry (MS) in clinical analysis dates back to the 1950s [2,3], when it was first employed to identify and quantify exogenous and endogenous compounds, such as drugs, metabolites, and proteins, in tissue samples. Technological advances have led to smaller instruments capable of providing detailed molecular information from tissue samples in real-time, with minimal sample pre-treatment. The MS technology can be used for classifying tissues and provides valuable prognostic information, such as the tumor subtype and grade [4]. MS technologies for tissue analysis now give promising results in precision medicine and intraoperative evaluation of surgical margins in various cancers, such as breast, pancreas, glioma, lung, brain, and liver [5,6,7,8,9,10,11]. These examples mostly employ an approach based on the identification of MS signals specific to tumors, which are significantly stronger than those of non-tumor tissues. The changes usually come from alterations to the cell metabolism or tumor microenvironment. MS-based strategies have been extensively investigated as prospective clinical tools for analyzing clinical materials such as tissues and biological fluids in clinical research, as well as for preoperative and intraoperative applications.

MS strategies used in clinical settings, especially those for the characterization of resection margins, fall largely into two classes: mass spectrometry imaging (MSI), where the tissues slices are used [12,13,14], and direct tissue sampling under ambient conditions.

While MSI is very useful for tissue phenotyping and classification, its implementation in the routine clinical and surgical setting is difficult, because of the time-consuming sample preparation, while direct tissue sampling MS can analyze samples under ambient conditions [15], with minimal or no sample pre-treatment. It is therefore suitable for routine work in the hospital. The great advantage of ambient ionization MS techniques is that they generate ions directly from the tissues in real time, with a minimal loss of components that are easily lost during sample pre-treatment [16]. Ease in operation on a real-time basis makes direct ambient MS extremely appealing, potentially suitable for daily clinical use.

Since its introduction into the operating room (OR) over a decade ago [17], two different techniques have been developed: on-line direct intraoperative MS and off-line sampling probe-based methods. On-line intraoperative mass spectrometry employs an authentic electrical knife, which causes thermal or mechanical disintegration of tissues. This leads to the generation of charged droplets, which, in turn, become gas-phase ions. The aerosol is collected through a vacuum pump and delivered to the inlet orifice of the mass spectrometer. The representative of this system is Rapid Evaporative Ionization Mass Spectrometry (REIMS), which first appeared in 2008 [5,18,19]. REIMS completes the whole analysis procedure in 3 s, with an extremely high specificity ranging from 92–100% for tissue identification. The MasSpec Pen device [20] also demonstrated highly promising results in the rapid diagnosis of clinical studies. Here, a small water flow is delivered through a pen-shaped device to biological samples. Biomolecules on the tissue surface are collected and analyzed by mass spectrometry in a few seconds. Off-line methods use a small tissue portion, which is directly analyzed by MS. Sampling these tissues with a Probe Electrospray Ionization (PESI) approach [21] was presented in clinical use 10 years ago by Takeda and colleagues [22,23] and promptly demonstrated the potential for direct tissue analysis [24,25,26,27].

The instrument analyses a very small piece of tissue that is homogenized and ionized by a needle probe at optimal conditions. Here, the needle probe is rapidly moved up and down along a vertical axis, penetrating the tissue sample. The needle, when pushed up to the highest point, after capturing the sample at the needle tip, generates an electrospray due to the high voltage applied to the needle probe, in front of the sampling orifice of the mass spectrometer. Based on the mass spectral profile from specific tissues (technically, the peak intensity ratios and mass values), the probability of either cancer or healthy tissue is displayed in a couple of minutes, expressed as the Tumor Cell Percentage (TCP).

Grouping is based on a Support Vector Machine (SVM), which is a supervised machine learning algorithm for generic pattern identification. For each disease, a specific dataset can be built and updated with new information for more accurate predictions.

In this work, we analyzed a large dataset of human tissue specimens, from Japan and Italy, with a PESI-MS instrument, to assess the reliability of the method in discriminating tumors from healthy tissues in different populations so that genetics, diseases, etiology, or operating room guidelines do not affect the instrument’s classification of specimens.

2. Materials and Methods

2.1. Japanese Patients’ Data and Specimen Collection

A total of 200 Japanese patients (

Table 1. Japanese and Italian specimens collected and analyzed by PESI-MS.

Tissue Specimens	Japanese Specimens	Italian Specimens
Neoplastic	103	118
Not neoplastic	97	106
Total	200	224

) were registered in this clinical study, whose entire protocol was approved by the IRB of each institution. The pathohistological diagnosis of each case included hepatocellular carcinoma (HCC), intrahepatic cholangiocellular carcinoma (ICC), cavernous hemangioma, hepatic adenoma, xanthogranuloma, and primary biliary cholangitis. While non-tumor tissues are considered to be free of cancer, most of the cases suffered from an underlying hepatic disease regarded as the cause of the malignant hepatic tumor.

Tissue specimens with a definitive pathology were collected from four high-volume centers in Japan: the University of Yamanashi Hospital (Chuo), the University of Tokyo Hospital (Tokyo), Japan Red Cross Medical Center (Tokyo), and National Cancer Institute Hospital (Tokyo). Small aliquots of specimens from the patients diagnosed with primary liver tumor were analyzed by PESI-MS, and their mass spectra were processed as previously described [25].

2.2. Italian Patients’ Clinical Data and Specimen Collection

Tumor and corresponding non-tumor tissue specimens were obtained, after authorization of the project by the competent research ethics and medico–ethical committee, as excess material from surgical operations in the Department of Hepatobiliary and General Surgery of the Humanitas Clinical and Research Center of Milan (Italy). Specimens were collected from a total of 224 patients who underwent liver resection for HCC (Table 1).

Inclusion criteria for the selection of patients were the provision of written informed consent and the availability of clinical, oncological, and pathological data. Patients whose data were missing were excluded. Specimens in the certified biobank were originally collected without affecting the diagnostic integrity of each sample, so the overall tissue was more than sufficient for histological diagnosis.

Tissue specimens from hepatocellular carcinoma and the corresponding non-tumor liver specimens were collected from 118 patients suffering from primary liver cancer. Once the surgical piece was removed from the patient, the surgeon placed it in a sterile container and delivered it to the pathological anatomy laboratory at room temperature. The pathologist selected a fragment of representative tissue, carefully excluding any necrotic regions and the corresponding portion of non-tumor tissue that was equally preserved. The selected tissue specimens were then divided into 50–100 mg (~0.5 cm³) pieces, immediately frozen in liquid nitrogen, and stored at −80 °C until analysis.

2.3. Chemicals and Reagents

Ethanol and 2-propanol were purchased from Carlo Erba (Cornaredo, Milano, Italy) at chromatographic grade. The water was purified using a Milli-Q system (Millipore, Bedford, MA, USA). Triol-type polypropylene glycol (PPGT, MW = 300, 700, 1500) standard solutions were obtained from Wako Chemicals (Richmond, VA, USA), and the PPGT mixture calibration solution was prepared in 2-propanol 50% and NaCl 5%.

2.4. PESI-MS Analysis

Analyses were carried out on a single quadrupole DPiMS-2020 mass spectrometer equipped with a PESI-MS/SVM Rapid Cancer Diagnostic Support System (Shimadzu Corporation, Kyoto, Japan), which used the PPGT mixture described as the standard sample for verifying instrument performance, checking peak intensity ratios and mass values, at the beginning of each day of analysis.

For each specimen, a 2 mm diameter piece of tissue was cut with a scalpel and homogenized with 100 µL of ethanol/water (50:50). Homogenization is performed in a 1.5 mL Eppendorf conical propylene microcentrifuge tube using a 0.5 mL pestle with a motorized pestle mixer (Argos Technology Life Science Products) for 10 s.

A total of 10 µL of the solution was dispensed in the solvent drip position of the sample plate and analyzed as described previously [24,28]. Briefly, for each tissue sample, the MS acquired the full scan spectra (in the m/z range 10–2000) for 1 min. For each acquisition, 10 equal fragments were created, and the spectra were averaged for each fragment. In this way, each tissue sample was characterized by 10 mass spectra that were used for statistical analysis in order to control errors or different sensitivities due to chemical noise or matrix effects. Ionization occurred on the probe needle of the PESI ion source, with an ionization voltage of −3.7 kV. The needle probe was replaced with a new one for each sample to prevent ionization loss and sample carryover.

Measurements were based on the dataset built on the results of 200 Japanese patients. The PESI-MS system classified the data obtained after ionization, based on a support vector machine (SVM) algorithm, and displayed the probability of the disease as the TCP. For a TCP ≥50%, specimens were classified as tumors, and TCP specimens <50% were classified as non-tumors.

2.5. Application of Artificial Intelligence

The DPiMS-2020 mass spectrometer available for the study was equipped with a dedicated “Statistical Analysis Software for Data from Direct Ionization Mass Spectrometry” software package (eMSTAT Solution^TM, software version 223-25864, October 2018, Shimadzu Corporation, Kyoto, Japan). This package allows a fully automatic statistical analysis of samples, multivariate analysis, and discriminant analysis to create AI models. The database was composed of both tumor and non-tumor tissues. The training data, which contained the data independent from the test specimens, was fed into the support vector machine (SVM) to train the learning machine. The training of the model was based on a pathologist’s assessment, used as the reference value for tissues’ classification. Histopathology was the basis of the HCC diagnosis and its differential diagnoses. The criteria employed here are discussed periodically by the WHO, and those currently used are reported in the 5th edition of the WHO Classification of Tumors of the Digestive Tract [29]. They include morphological features, such as cytological and architectural atypia, and immunohistochemical markers, such as GPC3, HSP70, and GS [30,31]. The approach has been described elsewhere [25]. Briefly, each MS spectrum was binned into intensity data for each 1 Da, normalized by its total ion current (TIC), and saved as a database for machine learning. The SVM model was then used for the binary classification of cancer and non-cancer. The kernel function used a quadratic expression, and the evaluation function in the model optimization was performed using the harmonic mean of sensitivity and specificity. The discriminant model was evaluated by using a 10-fold cross-validation.

3. Results

PESI-MS Analysis

On the 224 specimens collected by the Department of Hepatobiliary and General Surgery of Humanitas Clinical and Research Center of Milan, 311 PESI-MS analyses were conducted.

Although the majority of analyses agreed with the pathologist’s diagnosis, there were a few false positives (specimens defined as non-tumor by the pathologist but tumor by the instrument) and false negatives (specimens defined as tumor by the pathologist but non-tumor by the instrument). There were 266 analyses whose results were concordant with the pathologist’s diagnosis, 36 false positive, and 9 false negative (Figure 1). Considering the total number of analyses done by PESI-MS, the diagnostic concordance rate was 85.5% (n = 266).

Most of the non-concordant results were between probability values of TCP 40–70%. A set of these samples was selected randomly and re-analyzed. Those with a correct assignment confirmed the classification, while some of the borderline samples that had been misclassified the first time were then concordant with the pathologist’s diagnosis. Excluding non-concordant results that agreed with the pathologist on re-analyzing a second piece of specimens, the diagnostic concordance rate rose to 95.1% (Figure 2). On the samples that still resulted in being non-concordant, a follow-up was advisable, in order to check if the error came from the instrument or from the pathologist classification.

The false positive results were concentrated near the threshold of the HCC judgment score (50% TCP), and the concordance of the Italian specimens with the database created with the Japanese specimens was slightly low. This misclassification was supposed to be due to the higher sensitivity of mass spectrometry toward cancer cells compared to the pathologist’s subjective diagnosis. False negatives, on the other side, were all with low TCP, where less pathological tissue was present, and it was more difficult for the pathologist to diagnose it.

To improve the judgment concordance rate of SVM, a set of 30 new samples from Italian patients were added to the database created with Japanese samples (Figure 3). HCC judgment scores evaluated before and after additional training were compared: before the additional training, the peaks of the frequency distribution of HCC and non HCC specimens were around the threshold (Figure 3a), whereas after training, the peaks of the frequency were better separated (Figure 3b).

The performance of the database built by combining both Japanese and Italian samples was evaluated by the K-fold cross validation method, dividing the 2529 spectra acquired from Italian samples into K = 10 data sets (

Table 2. K-fold cross validation results both for Non-HCC and HCC spectra.

	Non-HCC			HCC			Total
Data Set No.	Number of Non-HCC Spectrum	Number of Correct Answer	Correct Answer Rate	Number of HCC Spectrum	Number of Correct Answer	Correct Answer Rate	Number of Total Spectra	Number of Correct Answer	Correct Answer Rate
Set 1	142	141	0.993	120	115	0.958	262	256	0.977
Set 2	107	107	1	152	144	0.947	259	251	0.969
Set 3	129	128	0.992	129	129	1	258	257	0.996
Set 4	120	115	0.958	132	108	0.818	252	223	0.885
Set 5	104	103	0.99	154	142	0.922	258	245	0.95
Set 6	130	125	0.962	131	131	1	261	256	0.981
Set 7	143	137	0.958	119	119	1	262	256	0.977
Set 8	120	117	0.975	143	138	0.965	263	255	0.97
Set 9	108	108	1	133	103	0.774	241	211	0.876
Set 10	105	98	0.933	108	108	1	213	206	0.967
Total	1208	1179	0.976	1321	1237	0.939	2529	2416	0.955

). The judgment concordance rate was 97.6% for non-HCC specimens and 93.9% for HCC specimens. Considering both HCC and non-HCC samples, the judgment concordance rate was 95.5% (see Table 2)

4. Discussion

The scope of this work was to verify that no influence from genetics, diseases, etiology, or operating room guidelines could affect the instrument in the classification of specimens. We tested the suitability of a specific HCC dataset, built for this system, by measuring 200 real specimens from Japanese patients, to correctly diagnose the tumor and non-tumor tissues of Italian patients. The Japanese database was also implemented with Italian samples, and this increased the model performance, providing a better separation of samples as SVM scores and a higher diagnostic accuracy, as compared with pathologist diagnosis.

The results provided in this study showed a high diagnostic concordance rate with the pathologist, demonstrating the versatility of the PESI-MS diagnostic system. As almost all of the false positive results were around a TCP = 50%, the discrepancies may have been due to the higher sensitivity of the instrument compared to the microscopic observation of the pathologist. The PESI-MS did in fact detect the presence of fewer tumor cells that are normally not visible to the pathologist with a classical diagnosis based on a microscopic histological analysis.

Finally, these results demonstrate the ability of the PESI-MS system to rapidly distinguish tumor and non-tumor tissue of Italian patients, based on the Japanese dataset, with no influence of racial or etiological characteristics or operating room procedures. Moreover, integrating the Japanese dataset with the mass spectra from the analysis of Italian samples increased model performance, increasing the ability of the instrument to classify tumors and non-tumor tissues specimens. Future studies are aimed at the identification of features that are important in specimens’ classification, through classical targeted metabolomics approaches, to provide fundamental evidence of tissue alterations due to these specific pathological alterations. The identification of precise discriminant features for HCC vs. non-HCC tissue, also, would allow the development of statistical approaches that do not require such extensive training, as we presented in this study.

5. Conclusions

In this study, the newly developed system to classify cancer tissues from non-tumor tissues proved to be able to carry out the classifications of tissues. For the first time, two different large ethnic cohorts were investigated to demonstrate the reliability and robustness of this approach.

It is worth noting that the specimens were classified using the pathologist’s diagnosis as the gold standard; obviously, there was no certified reference material to refer to. As a result, it is impossible to rule out the possibility that some system misclassifications, or completely inaccurate training, were due to issues in the pathologist’s classification of specific samples. Regarding Italian specimens, the pathology group was in a leadership position in the definition of both morphological and phenotypical diagnostic criteria [30,31]. This was possible not only due to the specific expertise in histopathology, but also due to a large amount of surgical and biopsy specimens available in the hospital biobanks. In Japan, the same approach was used. Despite representing the gold standard in diagnosis, pathology might be affected by errors, ranging from 0.1% to 0.8% in a large series of surgical specimens [32,33]. It might be useful, in future studies, to consider the pathologist’s method of detecting HCC, possibly using standardized procedures, at least for the model training process, or using, as a reference, the classification of the same tissue given by two or more pathologists.

It is also worth noting that the misclassifications found were almost all clustered around the tumor/non-tumor discrimination point, at a region where the sample’s histology was borderline.

Because the instrument has been also used in a biosafety level 3 (BL3) environment, for this reason, it was subjected to peroxidation after it was removed.

Despite being conceptually ready for the logistics aspects of entering the clinic and being used by the pathologist as a support for diagnosis, there are still numerous obstacles to overcome. In addition to fundamental research aspects, there are cultural aspects, such as the acceptability of automatic diagnostics, and regulatory aspects, such as those that regulate medical diagnostic equipment that requires a complete device, methods, and consumables certifications, which are hoped to be developed as soon as possible for this promising new tool.