Unsupervised feature dimension reduction for classification of MR spectra.

We present an unsupervised feature dimension reduction method for the classification of magnetic resonance spectra. The technique preserves spectral information, important for disease profiling. We propose to use this technique as a preprocessing step for computationally demanding wrapper-based feature subset selection. We show that the classification accuracy on an independent test set can be sustained while achieving considerable feature reduction. Our method is applicable to other classification techniques, such as neural networks, support vector machines, etc.

Diagnosis and prognosis of breast cancer by magnetic resonance spectroscopy of fine-needle aspirates analysed using a statistical classification strategy.

BACKGROUND: The aim was to develop robust classifiers to analyse magnetic resonance spectroscopy (MRS) data of fine-needle aspirates taken from breast tumours. The resulting data could provide computerized, classification-based diagnosis and prognostic indicators. METHODS: Fine-needle aspirate biopsies obtained at the time of surgery for both benign and malignant breast diseases were analysed by one-dimensional proton MRS at 8.5 Tesla. Diagnostic correlation was performed between the spectra and standard pathology reports, including the presence of vascular invasion by the primary cancer and involvement of the excised axillary lymph nodes. RESULTS: Malignant tissue was distinguished from benign lesions with an overall accuracy of 93 per cent. From the same spectra, lymph node involvement was predicted with an overall accuracy of 95 per cent, and tumour vascular invasion with an overall accuracy of 94 per cent. CONCLUSION: The pathology, nodal involvement and tumour vascular invasion were predicted by computerized statistical classification of the proton MRS spectrum from a fine-needle aspirate biopsy taken from the primary breast lesion.

Cryptococcomas distinguished from gliomas with MR spectroscopy: an experimental rat and cell culture study.

PURPOSE: To use magnetic resonance (MR) spectroscopy to characterize clinical isolates of Cryptococcus neoformans and a glioma cell line in culture and in experimental rats. MATERIALS AND METHODS: One- and two-dimensional hydrogen 1 MR spectra were acquired from fungi cultured in vitro (16 isolates of C neoformans, three of Candida albicans, three of Aspergillus fumigatus, three of Saccharomyces cerevisiae) and a C6 glioma cell line. Cerebral biopsy specimens were obtained from healthy rats and animals with experimental infections or gliomas (19 healthy brains, 20 cryptococcomas, and 19 gliomas). Unequivocal signal assignment was performed for cell suspensions and tissue samples by using homo- and heteronuclear two-dimensional correlation spectra. RESULTS: MR spectra of C neoformans and cerebral cryptococcomas--but not of other fungi, healthy brains, or gliomas--were dominated by resonances from the cytosolic disaccharide alpha,alpha-trehalose. This spectral pattern was different from that of gliomas, which was dominated by lipids and an increased choline-creatine ratio, and that of healthy brain. CONCLUSION: A remarkably high concentration of alpha,alpha-trehalose in relation to other metabolites that are visible with MR spectroscopy is diagnostic of C neoformans. Cerebral cryptococcomas are an uncommon but serious manifestation of cryptococcosis in humans. Application of these results to the noninvasive diagnosis of cerebral cryptococcomas would help reduce the risk and expense of unnecessary surgery or biopsy and expedite patient treatment.

Fine-needle biopsy specimens of benign breast lesions distinguished from invasive cancer ex vivo with proton MR spectroscopy.

PURPOSE: To determine whether invasive breast cancer can be distinguished from benign lesions with proton magnetic resonance (MR) spectroscopy ex vivo on the basis of altered cellular chemistry. MATERIALS AND METHODS: Two hundred eighteen fine-needle biopsy specimens were obtained in 191 patients undergoing surgery and were analyzed with proton MR spectroscopy. MR spectroscopic and histopathologic findings were compared. RESULTS: Invasive carcinoma produced increased signal at 3.25 ppm, attributable to choline-containing metabolites. Discrimination between invasive carcinoma (n = 82), benign lesions (n = 106), or carcinoma in situ (n = 17) was based on the resonance intensity at 3.25 ppm standardized to the resonance at 3.05 ppm (P < .001). The ratio of peak height intensities of resonances at 3.25 to those at 3.05 ppm was less than 1.7 in 102 of the 106 normal or benign lesions. All carcinoma in situ specimens with comedonecrosis or a microinvasive component (n = 6) were categorized at MR spectroscopy with invasive carcinoma, while others with in situ disease alone were categorized with benign lesions (n = 11). The sensitivity and specificity of MR spectroscopy in fine-needle biopsy specimens in distinguishing benign lesions from invasive cancer were 95% and 96%, respectively. CONCLUSION: Proton MR spectroscopy of fine-needle biopsy specimens provides objective diagnostic information that complements findings of conventional preoperative investigations of breast lesions.

Cancer pathology in the year 2000.

The last one hundred and fifty years has produced the mature and sophisticated discipline of histopathology, yet still leaves the diagnosis of human cancer, by the best available technique, as more art than science. Proton magnetic resonance spectroscopy (1H MRS) ex vivo identifies the chemical markers of established pathobiological disorders within excised biopsies and fine needle aspirates, in particular, those associated with the development and progression of malignant disease. Alterations to cellular chemistry monitored by 1H MRS allows distinction between invasive and pre-invasive lesions of the uterine cervix, and separate truly benign follicular neoplasms from follicular carcinomas on analysis of fine needle aspirates containing as few as 10(6) cells. 1H chemical shift imaging (CSI) determines the spatial location of these chemical changes and provides insight into the chemistry of neoplastic transformation. It is our hypothesis that, by the year 2000, CSI will aid image guided biopsy techniques and that correlation of biopsy histology with in vivo localised 1H MRS data will: (a) lead to improved assessment of the extent of malignant disease and (b) establish the sensitivity and specificity of in vivo 1H MRS for the simultaneous determination of the size, location and neoplastic potential of a tumour mass.

A proton magnetic resonance spectroscopy study of aging and transformed human fibroblasts.

Proton magnetic resonance spectroscopy (1H MRS) has been used to monitor changes occurring during aging and transformation in human lung fibroblasts. Aging was studied in MRC-5 cells from nonsenescent (early passage) to presenescent (late passage) and senescence. Nonsenescent cells infected with SV40 virus (pretransformed) were monitored through crisis and subsequent immortalization. Aging changes were observed with one- and two-dimensional MR spectra. Cholesterol and lipid resonances were significantly increased from nonsenescent cultures to senescence. These changes could be caused by chemical or structural changes in the plasma membrane or in intracellular lipid pools. In contrast, choline levels rose from nonsenescent to presenescent cells but at senescence dropped to that of nonsenescent cells. Increased choline levels are often associated with increased cellular proliferation. After SV40 infection of MRC-5 cells there was an increase of cholesterol and lipid levels that peaked at crisis. Newly immortalized cells exhibited a drop in cholesterol and lipid to nonsenescent cell levels, but these rose again in established immortalized cells. In contrast to presensescent cultures, the levels of choline gradually increased from pretransformed to crisis phase but still continued to rise after immortalization. Thus, 1H MRS illustrates similarities in lipid behavior at senescence and crisis, whereas the choline levels are different.

Two-dimensional proton magnetic resonance spectroscopy for tissue characterization of thyroid neoplasms.

We have previously demonstrated that one dimensional (1D) proton (1H) magnetic resonance spectroscopy (MRS) can distinguish normal thyroid tissue from thyroid carcinoma using a spectral ratio of peak intensity at 1.7 ppm/0.9 ppm. Two dimensional (2D) 1H-MRS allows identification of specific molecules that have overlapping peaks in the 1D-MR spectrum. Specimens from 93 consecutive thyroid nodules were examined using 2D 1H-MRS on a Bruker AM-360 wide-bore spectrometer. There was a progressive increase in lipid cross peaks assigned to di-/triglycerides when comparing colloid/hyperplastic nodules to follicular adenoma, and adenoma to carcinoma. A specific cross peak attributable to cholesterol/cholesteryl esters was commonly seen in carcinomas. In contrast, two unassigned cross peaks unique to the thyroid were more prevalent in benign lesions. There was an overall increase in cross peaks attributable to cell surface fucosylation in carcinoma when compared to benign lesions, although the fucose spectral pattern was not specific for cancer. On this basis, a spectral ratio of peak intensity at 2.05 ppm/0.9 ppm more clearly distinguished benign follicular adenoma from carcinoma. 2D 1H-MRS thus identifies chemical changes that allow more specific tissue characterization of thyroid neoplasms.

Tetraphenylphosphonium chloride induced MR-visible lipid accumulation in a malignant human breast cell line.

The effect of the cationic lipophilic phosphonium salt tetraphenylphosphonium chloride (TPP) on a human malignant breast cell line, DU4475, was monitored with proton nuclear magnetic resonance (1H MRS). TPP caused a dose- and time- dependent increase in resonances arising from MR-visible lipid as measured by the CH2/CH3 ratio in the 1-dimensional 1H MR spectrum. Two-dimensional MRS identified increases in the glycerophosphocholine/lysine cross-peak ratio and corresponding decreases in the phosphocholine/lysine ratio in a dose- dependent fashion in TPP-treated cells. Lipid metabolic changes are discussed in the light of other MR experiments, and the data indicate that accumulation of MR-visible lipids may arise from the rearrangement of phospholipids accompanying mitochondrial destruction or from the catabolism of phospholipids associated with early events in the cytotoxic process.

Chemical shift imaging of human colorectal tissue (ex vivo).

The spatial location of MR visible lipid in the wall of the normal human colon, and in carcinomatous colonic tissue has been documented using proton chemical shift imaging, one- and two-dimensional magnetic resonance spectroscopy and histochemical staining. Following dissection of the mucosal and submucosal layers of normal colon, these techniques showed high levels of neutral lipid distributed in the submucosal layer. Relatively less lipid was observed in the mucosal layer. Histochemical staining confirmed that the majority of the neutral lipid was in the submucosa, extracellular, and in the lymphatic channels. Carcinomatous tissue gave a variable lipid signal which histochemical staining identified as being from tumour stroma, necrotic and degenerate tumour cells and macrophages.

Human cancers detected by proton MRS and chemical shift imaging ex vivo.

Proton magnetic resonance spectroscopy (1H MRS) has the potential to become a diagnostic adjunct for the detection and grading of human neoplastic disease. This paper describes the use of proton MRS to document changes arising in the lipid chemistry of biopsies arising from the human uterine cervix, thyroid and colon and demonstrates the diagnostic power of ex vivo spectroscopy. Proton chemical shift imaging (CSI) is further used to determine the spatial location of lipid changes in ex vivo human biopsy specimens and provides insight into the chemistry of neoplastic transformation.

Proton magnetic resonance and human thyroid neoplasia III. Ex vivo chemical-shift microimaging.

Magnetic-resonance chemical-shift microimaging, with a spatial resolution of 40 x 40 microns, is a modality which can detect alterations to cellular chemistry and hence markers of pathological processes in human tissue ex vivo. This technique was used as a chemical microscope to assess follicular thyroid neoplasms, lesions which are unsatisfactorily investigated using standard histopathological techniques or water-based magnetic-resonance imaging. The chemical-shift images at the methyl frequency (0.9 ppm) identify chemical heterogeneity in follicular tumors which are histologically homogeneous. The observed changes to cellular chemistry, detectable in foci of approximately 100 cells or less, support the existence of a preinvasive state hitherto unidentified by current pathological techniques.

Modulation of MR-visible mobile lipid levels by cell culture conditions and correlations with chemotactic response.

A transformed murine fibroblast cell line has been used to assess which criteria govern the appearance of a lipid pool that is mobile on the MR time scale. A high-resolution proton MR signal arising from neutral lipids, including triglyceride and cholesteryl esters, has previously been associated with membrane events in stimulated, transformed and malignant cells. We report that the attenuation of cellular proliferation by confluence or low pH caused significant increases in MR-visible lipid and that the lipid signal could be amplified at high density by the removal of serum. A significant decrease in chemotactic response accompanied the culture of cells at high density, but chemotactic response was not generally linked to alteration of the lipid signal. The appearance of the signal was also not correlated with the proportion of cells in any phase of the cell cycle. Significant changes in the MR-visible pools of the lipid metabolites choline, phosphocholine and glycerophosphocholine were measured under the culture conditions employed with 2D MRS and suggest that MR-visible lipid may arise from the catabolism of phospholipids.

Characterization of human ovarian epithelial tumors (ex vivo) by proton magnetic resonance spectroscopy.

Proton magnetic resonance spectroscopy (1H MRS) offers an alternative investigational modality that will assist current pathologic techniques in the diagnosis of human ovarian epithelial tumors. Histologically normal human ovarian tissue (n = 12) was compared with ovarian benign fibromas (n = 3) and surface epithelial-stromal tumors (benign, n = 18; proliferating, n = 9; frankly malignant, n = 30) ex vivo by 1H MRS. The distinction between carcinomatous and benign or normal tissue (P<0.0001; Student's t-test) was made on one-dimensional (1-D) 1H MR spectra utilizing differences in resonance intensities of cellular lipid, creatine/phosphocreatine and lysine. The sensitivity and specificity of the test were 87% and 91%, respectively. Two-dimensional (2-D) MRS of carcinomatous biopsies showed multiple crosspeaks attributable to cell-surface fucosylation that correlated with tumor grade and loss of cellular differentiation. The multiple fucose crosspeaks were absent in spectra from normal ovary and benign tumors. The distinction between carcinomatous and normal or benign tissue based on MR-visible fucosylation gave a sensitivity and specificity of 88% and 97%, respectively. Proliferating tumors exhibited a range of cell-surface fucosylation patterns indicative of malignant potential.

Diagnosis of follicular thyroid lesions by proton magnetic resonance on fine needle biopsy.

Most thyroidectomies are currently performed for diagnostic purposes. It has been established that proton magnetic resonance spectroscopy (MRS) on excised thyroid tissue can distinguish normal thyroid from invasive carcinomas (P < 0.0001). The purpose of this study was to assess whether the same discrimination could be obtained preoperatively from fine needle biopsy (FNB). This has clinical importance because cytological examination of fine needle aspirates cannot distinguish between benign and malignant follicular thyroid lesions. Here we demonstrate a sensitivity of 95% for proton MRS to correctly identify clinically or histologically proven carcinoma. MRS measurements were made on FNB specimens (containing as few as 10(6) cells) from solitary thyroid nodules. MR assessment of FNB was inconsistent with that of the corresponding tissue in only 6.5% of cases. The discrimination between cancer and normal tissue was based on altered cellular chemistry measured as a one-dimensional spectral ratio of resonances from the amino acid lysine and lipid. Benign follicular lesions were separated into two groups: 67% with a spectral ratio similar to malignant thyroid tumors, and 33% with a spectral ratio comparable to that in normal thyroid tissue. Thus, in contrast with histopathology, MRS offers a method for assessment of FNB of follicular lesions with the potential to identify a biologically benign group, which could avoid thyroid surgery for purely diagnostic purposes.

Correlation of cellular differentiation in human colorectal carcinoma and adenoma cell lines with metabolite profiles determined by 1H magnetic resonance spectroscopy.

The aim was to determine whether proton magnetic resonance spectroscopy (MRS) could grade human colorectal cells of differing malignant potential. A cell model of tumour development and progression comprising 2 non-tumorigenic adenoma lines and 4 carcinoma lines of increasing tumorigenicity was chosen. A gradual reduction in cellular differentiation and an accumulation of genetic alterations from adenoma to carcinoma characterized the selected cell lines. One-dimensional and 2-dimensional MRS showed that reduced differentiation in the cell model correlated with an increase in the levels of lipid, metabolites, the glycosylation intermediate uridine diphospho-N-acetylglucosamine and cell-surface fucosylation. Mutations involving the K-ras, APC and DCC genes are present both in adenoma- and in carcinoma-derived lines in this model, but the first evidence of an abnormality in the p53 gene was concomitant with the cells' ability to grow as a tumour in athymic nude mice. This genetic change coincided with the detection, by MRS, of UDP-hexose (ribose moiety, 2D MRS cross peak between H2 at 4.38 ppm and HI at 5.99 ppm) and the appearance of an additional fucosyl resonance (cross peak between-CH3 at 1.41 and H5 at 4.30 ppm) in the least tumorigenic of the carcinoma cell lines. An increase in complexity of the fucosylation spectral pattern was observed with further cellular de-differentiation and increased tumorigenicity. Collectively these data support the existence of an adenoma-carcinoma sequence.

Proton magnetic resonance and human thyroid neoplasia. II: Potential avoidance of surgery for benign follicular neoplasms.

Thyroid cancer is rare, but many thyroidectomies continue to be performed simply to exclude a diagnosis of malignancy. The purpose of this study was to determine the potential financial savings associated with the use of proton magnetic resonance analysis of follicular neoplasms. Proton magnetic resonance spectroscopy was performed on tissue obtained at the time of surgery from 98 consecutive solitary or dominant thyroid nodules. Fine-needle biopsies were also performed on operative specimens, and the tissues assessed by proton magnetic resonance; these spectra were then compared with those obtained from tissue specimens. An estimate of potential savings was obtained by comparing the magnetic resonance data with the indications for surgery and pathology on all patients having thyroidectomy over a 10-year period. Proton magnetic resonance spectroscopy was able to distinguish between normal thyroid tissue and invasive thyroid cancer with 100% specificity. Benign follicular adenomas fall into two groups: 44% having a spectral pattern comparable with normal thyroid, and the remaining 56% demonstrating an altered spectral pattern more comparable to the malignant magnetic resonance profile. Proton magnetic resonance spectroscopy on fine-needle biopsy specimens produced spectra similar to those from tissues from the same patient. From a fine-needle biopsy specimen, proton magnetic resonance spectroscopy can identify a group of benign follicular adenomas with spectral profiles akin to those of normal thyroid cells, thus avoiding the need for unnecessary surgical excision. The potential savings in one surgical unit alone were over $1 million in 10 years.

Proton magnetic resonance and human cervical neoplasia. II. Ex vivo chemical-shift microimaging.

Proton chemical-shift imaging at 8.5 T has been used to detect malignant foci in small (6 mm3) biopsies from the human uterine cervix. Images based on the lipid resonances of frankly malignant cells discriminate between tumor tissue and host stroma and distinguish invasive from preinvasive cervical cancer (n = 7). With this method, foci of malignant cells were revealed in 500 micron slices with an in-plane resolution of 40 by 160 microns. The MR intensity maps reflected the local distribution of malignant cells as assessed by histopathology. The lower signal-to-noise ratio inherent for these non-water-based images was improved by applying postacquisitional matched Gaussian window functions, thus effecting a substantial increase in contrast with minimal loss in spatial resolution.

Proton magnetic resonance spectroscopy of lymphocytes: an historical perspective.

Proton magnetic resonance spectroscopy can be used to elucidate alterations to cellular chemistry associated with specific biological functions. It became apparent that this type of information was present in the magnetic resonance spectra from intact viable lymphocytes in the late 1970s. It was not until the 1980s, however, that one-dimensional multipulse sequences were used to filter the large signal contributions from water and fat, which had until then masked weaker signals from other molecules. When this technology was combined with two-dimensional spectroscopy, unambiguous assignment of the biologically relevant chemical species became possible. In vitro activated, stimulated, transformed, and malignant lymphocytes, as well as embryonic fibroblasts and malignant cells of epithelial origin, all gave rise to a strong triglyceride spectrum and resonances from a multitude of cellular metabolites. Two-dimensional spectroscopy and the analyses of highly purified membranes determined that the triglyceride signals originated, at least in part, from the plasma membrane. Based on physiochemical data, a new model for the structural arrangement of plasma membrane lipid in these cells was proposed. While differences exist between the proton magnetic resonance spectra of stimulated lymphocytes and malignant cells in vitro, they share a high-resolution lipid spectrum. In tissue, however, the presence of activated lymphocytes does not always produce the lipid spectrum, particularly in the vicinity of tumors.

Proton magnetic resonance and human thyroid neoplasia. I: Discrimination between benign and malignant neoplasms.

PURPOSE: Thyroid nodules are very common, yet the vast majority are biologically benign. The extreme difficulty facing the clinician selecting potentially malignant thyroid nodules for surgery was the subject of a recent editorial by Ernest L. Mazzaferri in the American Journal of Medicine (93:359-362, 1992). Here we evaluate the potential of proton magnetic resonance spectroscopy (1H MRS) to provide a solution to this problem. PATIENTS: Thyroid tissue from fifty-three patients undergoing partial or total thyroidectomy for solitary thyroid nodules were assessed by 1H MRS. RESULTS: When compared with the histologic diagnosis, 1H MRS distinguished normal thyroid tissue (n = 8) from invasive papillary (n = 9), anaplastic (n = 1), and medullary (n = 1) carcinomas with P values of < 0.0001, based on altered cellular chemistry. The same magnetic resonance (MR) criteria categorized pathologically proven follicular carcinoma (n = 8) (established as such by the presence of capsular or vascular invasion at the periphery of the tumor, or by the presence of metastases in the patient) with the other thyroid cancers (P < 0.0001). All other "benign" follicular neoplasms (n = 34), including five atypical follicular adenomas, were assessed by the same 1H MRS criteria and found to fit into one of the two above categories, viz. analogous to benign or malignant thyroid tissue. CONCLUSIONS: Proton MRS has the potential to separate out a group of truly benign follicular neoplasms from follicular tumors (both follicular adenomas and follicular carcinomas) that have an atypical follicular pattern on cytologic examination. This is the first report of an objective diagnostic procedure that has the potential to obviate surgical excision in a significant number of patients with benign follicular adenomas, independent of exhaustive histopathologic assessment.

Assessment of human colorectal biopsies by 1H MRS: correlation with histopathology.

Samples (3 mm3) of histopathologically normal (n = 15) and carcinomatous tissue (n = 15) were obtained from colectomy specimens and examined by 1H MRS. A combination of one- and two-dimensional spectra, obtained with appropriate acquisition and processing parameters, provide multiple diagnostic parameters allowing the distinction between normal and carcinomatous tissue. The diagnostic parameters include resonances from choline, choline-based, and other metabolites, cell surface fucosylation, and altered lipid profiles. Tissues histopathologically classified as normal, while remaining distinct from the malignant spectral profile, were found to fit into two categories, one of which had some of the spectral characteristics of malignancy. These results indicate that 1H MRS identifies abnormal colorectal mucosa, which is not morphologically manifest. Such abnormalities have been reported previously to exist in premalignant colorectal tissue by monoclonal antibody studies. Collectively, these results suggest that a clinical study of colorectal biopsies by 1H MRS could provide support for the use of MRS as an adjunct to current pathological procedures.